A method for transferring industrial operation skills to embodied robots

Abstract

This invention relates to the field of industrial robot control technology and discloses a method for transferring and training industrial operation skills for embodied robots. This method addresses the problems in traditional methods where inaccurate physical contact modeling and spatiotemporal asynchrony of multimodal perception easily lead to control mismatch at the moment of contact, resulting in end-effector flutter, contact force exceeding limits, workpiece damage, and equipment downtime. The method first performs unified time stamping and spatiotemporal alignment on visual data, force data, and robot body state data. It then adjusts the fusion weights of visual guidance and force feedback based on the degree of proximity. After contact is established, it adjusts contact control parameters based on the difference between the actual and simulated contact states. Finally, it combines joint transmission dead zone compensation, limit-exceeding retreat, state reset, and degraded retry for closed-loop control, achieving improved contact stability, production line safety, and processing yield in precision contact operations.

Description

Technical Field

[0001] This invention relates to the field of industrial robot control technology, specifically to a method for transferring and training industrial operation skills for embodied robots. Background Technology

[0002] Training robots' industrial operation skills in a 3D simulation environment before transferring them to a real production line is an important application direction of embodied intelligence in the field of intelligent manufacturing. Existing technologies have already conducted relevant research in this area. For example, the published invention patent application CN115990891B discloses a robot reinforcement learning assembly method based on visual teaching and virtual-real transfer, which improves the assembly success rate after strategy transfer through visual teaching, virtual environment training, domain randomization, and visual error estimation. Another example is the published invention patent application CN113134839B, which discloses a robot precision flexible assembly method based on visual and force-position image learning, which guides the robot to complete precision flexible assembly by learning visual images and force-position relationship images during the assembly process. While the aforementioned solutions improve robot assembly capabilities from the perspectives of virtual-to-real transfer training and assembly process learning, they primarily focus on assembly trajectory generation, improving transfer success rates, or assembly process control, failing to fully address contact safety issues in precision contact operations on real production lines. This is mainly because existing 3D physical simulations struggle to accurately characterize the rigid-flexible coupling contact dynamics under minute tolerance conditions; dynamic friction changes, contact collision responses, and joint transmission backlash still differ from real-world operating conditions. Furthermore, multimodal perception data, including visual, force, and body state perception, suffers from inconsistent sampling frequencies during real-world deployment. Issues such as inconsistency, inconsistent transmission delays, and difficulties in time alignment exist. Due to these factors, control strategies based on simulation prior training are prone to mismatch between perceived information and control parameters when performing precision contact operations such as compliant assembly, wire harness insertion / removal, or grinding, at the moment the robot comes into contact with the real workpiece. This can lead to high-frequency chattering in the end effector, excessive instantaneous contact force, damage to the precision workpiece, and even joint overload shutdown. Therefore, how to reduce the risk of control mismatch at the moment of contact during virtual-to-real migration, avoid damage to precision workpieces, and ensure stable operation of the production line has become an urgent technical problem to be solved in this field. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a method for transferring and training industrial operation skills for embodied robots. This method solves the problems in traditional methods, such as inaccurate physical contact modeling and spatiotemporal asynchrony of multimodal perception, which can easily lead to control mismatch at the moment of contact, resulting in end effector flutter, contact force exceeding limits, workpiece damage, and equipment downtime.

[0004] To achieve the above objectives, the present invention provides the following technical solution: A method for transferring industrial operation skills to embodied robots includes: S1. Obtain the baseline control strategy formed by simulation training, and uniformly time-stamp the visual data, force data and robot body state data to establish a historical state sequence. S2. Based on the acquisition time of the visual data, determine the corresponding robot body state from the historical state sequence, and perform spatiotemporal alignment processing on the visual data. S3. Adjust the fusion weight of visual guidance information and force feedback information according to the proximity between the end effector and the target contact area; S4. After detecting that the end effector has made contact with the target contact area, adjust the contact control parameters of the contact direction according to the difference between the actual contact state and the simulated contact state. S5. After detecting the joint transmission dead zone, apply a small-amplitude alternating compensation signal to the corresponding joint. S6. After the contact force exceeds the safety boundary, the control end effector retracts in the opposite direction and performs a state reset and downgrade retry.

[0005] Preferably, S1 includes: Complete the strategy file matching and load the baseline control strategy according to the task identifier, end tool coordinate parameters and workpiece type identifier; Using the controller master clock or fieldbus synchronous clock as a unified reference time source, the acquisition time is written to the acquisition data that has the ability to write local timestamps, and the arrival time is written to the acquisition data that does not have the ability to write local timestamps, and the link fixed delay correction is performed. Force data and robot body state data are recorded in chronological order using a circular buffer. Based on the comparison between the time interval between adjacent frames and the rated sampling period, abnormal data source identification, clock resynchronization, continuous monitoring, and access recovery are performed.

[0006] Preferably, S2 includes: Read the acquisition time label of the visual data, and use nearest neighbor matching and adjacent state interpolation in the historical state sequence to determine the corresponding robot body state; By combining the robot calibration model, hand-eye calibration matrix and visual acquisition device installation transformation relationship, spatial mapping and pose advance are performed on the point cloud data or depth-restored image contour data of the target area; The synchronization state is formed by marking the time matching status, the confidence level of target area recognition, and the cumulative delay of visual data.

[0007] Preferably, S3 includes: Determine the real-time approach distance based on the spatial location of the target contact area and the current end position; When the target contact area is a hole, groove or curved surface, the shortest distance from the center point of the end tool to the reference surface or reference center is used as the approach degree index, and the approach process is divided into macro approach interval, transition interval and micro contact preparation interval. The visual guidance weight and force feedback weight are continuously adjusted according to the smoothing rule; In the micro-contact preparation zone, visual guidance weight is reduced while the proportion of force feedback and end-effector micro-motion control is increased. Modal state marking is performed by combining force pre-triggering conditions and visual-force conflict determination to generate fusion weight reorganization and contact preparation state information.

[0008] Preferably, in the micro-contact preparation zone, the weight of visual guidance is reduced while the proportion of force feedback and control of the end effector's minute movement states is increased, including: When the end effector enters the micro-contact preparation zone but has not yet detected clear contact, the visual guidance weight is limited to a preset low value range to preserve the directional guidance effect of visual information. Force feedback information and the micro-motion state of the end effector are used as the main control basis, and the contact preparation state is determined based on the micro-positional change of the end effector, the rate of change of the end effector position, and the increase of the contact force.

[0009] Preferably, S4 includes: Contact confirmation is achieved by combining the rate of change of contact force, normal contact force, and the relative positional relationship between the end and the target surface; Match the reference contact trajectory according to the target area number, end attitude, approach velocity, and contact depth; Differences are classified based on contact force deviation, velocity attenuation deviation, displacement deviation, and jamming state; Adjust the stiffness and damping parameters in the contact direction according to the classification results; The system monitors and determines the contact oscillation state and the stable contact establishment state, and outputs a flutter risk indicator or a backoff request indicator.

[0010] Preferably, the differences are graded based on contact force deviation, velocity attenuation deviation, displacement deviation, and jamming state, including: The difference evaluation results are constructed using a rule-based hierarchical approach; The ratio of the actual contact force to the reference contact force is used to characterize the contact force deviation. The ratio of the actual approach speed to the reference speed attenuation is used to characterize the speed attenuation deviation. The difference between the actual displacement of the end along the contact direction and the reference contact displacement is used to characterize the displacement deviation; Insufficient displacement growth after contact is established and continuous accumulation of contact force are used to characterize the stuck state. Based on the above criteria, the conditions are classified as normal, first-level out-of-tolerance, and second-level out-of-tolerance.

[0011] Preferably, S5 includes: Based on the displacement difference or end micro-displacement calculation results between the joint motor side and the output side, the transmission dead zone is identified by combining the pre-calibrated static friction identification threshold, joint drive current, command direction, equivalent angular displacement difference, output side angular velocity and duration. After the determination is established, an alternating compensation signal is superimposed on the corresponding joint. The frequency and amplitude of the alternating compensation signal are determined based on the structural modal test results and static friction identification results. The determination of dead zone release, compensation cancellation control, or compensation failure is made based on the difference in equivalent angular displacement after compensation, the angular velocity on the output side, the continuity of micro-displacement at the end, the change in contact force, and the change in current within the observation window.

[0012] Preferably, S6 includes: When the contact force exceeds the process safety boundary corresponding to the current task, or the dead zone compensation fails and the contact force continues to increase, or the risk of continuous chatter is not eliminated after local parameter adjustment, stop advancing to the current position and retreat in the opposite direction of the current contact force vector; After the retreat is completed, clear the contact data fragments before the limit is exceeded, reconstruct the spatial relationship between the workpiece and the end, trigger the target area to be rescanned and reset the control parameters; When the target area offset is within the allowable range and the number of automatic retries has not reached the upper limit, the conservative retry parameter group is invoked. When the target area offset exceeds the allowable range or the number of automatic retries reaches the limit, the system enters the manual intervention state and outputs a process exception code.

[0013] Preferably, after the retreat is completed, the contact data fragments before the limit was exceeded are cleared, the spatial relationship between the workpiece and the end effector is reconstructed, the target area is rescanned, and the control parameters are reset, including: After the retreat ends, the state is reset, the out-of-limit contact data fragments in the historical state buffer are cleared, the spatial relationship between the current workpiece position and the end position is re-established, the target area scan is triggered, and the contact control parameters, fusion weights and compensation signal states are restored to the safe initial settings. The synchronization state before the limit was exceeded, the contact difference evaluation results, the process of fusion weight change, the joint dead zone identification results, and the contact force direction at the moment of retreat triggering were traced back, and abnormal samples were recorded.

[0014] Compared with existing technologies, this invention provides a method for transferring and training industrial operation skills for embodied robots, which has the following beneficial effects: 1. This invention establishes a unified time reference for visual data, force data, and robot body state data, performs spatiotemporal alignment processing on visual observation results, and combines continuous adjustment of multimodal weights during the approach phase, online adjustment of contact control parameters during the contact phase, joint transmission dead zone compensation, and reverse retreat, state reset, and degraded retry after exceeding limits. This enables the control strategy formed by simulation training to be corrected online for contact dynamic differences and multimodal timing asynchrony problems in real production lines. It reduces the risk of high-frequency flutter, excessive contact force, damage to precision workpieces, and joint overload shutdown caused by mismatch between instantaneous sensing information and control parameters during contact, thereby improving the contact stability, production line operation safety, and processing yield of the embodied robot in precision contact operations.

[0015] 2. This invention compares the actual contact state with the simulated contact state after contact is established, and combines joint transmission dead zone identification, alternating compensation, and abnormal sample backtracking to make targeted adjustments to the control process after the virtual-to-real migration. This enables the robot to maintain the continuity and adjustability of the operation process when facing workpiece tolerance fluctuations, local interference, transmission backlash, and changes in contact state in the real production line. It reduces repeated parameter adjustments, operation interruptions, and manual intervention caused by inconsistencies between simulation priors and on-site working conditions, and improves the deployment adaptability, operation continuity, and engineering application stability when industrial operation skills are migrated from the simulation environment to the real production line. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of a method for transferring and training industrial operation skills for embodied robots according to the present invention. Figure 2 This is a schematic diagram of the unified time stamp for multimodal data according to the present invention; Figure 3 This is a schematic diagram of the spatiotemporal alignment and pose advance of visual data in this invention; Figure 4 This is a graph showing the weight adjustment of proximity distance partitioning and visual-mechanical fusion in this invention. Figure 5 This is a flowchart of the contact difference evaluation and contact control parameter correction process of the present invention; Figure 6 This is a schematic diagram of the joint transmission dead zone identification and micro-amplitude alternating compensation of the present invention; Figure 7 This is the state machine diagram for the safety over-limit retreat and degradation retry of the present invention; Figure 8 This is a schematic diagram of the actual production line migration training system of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1: Figures 1-7 A method for transferring industrial operation skills to embodied robots is presented, including: S1. Obtain the baseline control strategy formed by simulation training, and uniformly time-stamp the visual data, force data and robot body state data to establish a historical state sequence. S2. Based on the acquisition time of the visual data, determine the corresponding robot body state from the historical state sequence, and perform spatiotemporal alignment processing on the visual data. S3. Adjust the fusion weight of visual guidance information and force feedback information according to the proximity between the end effector and the target contact area; S4. After detecting that the end effector has made contact with the target contact area, adjust the contact control parameters of the contact direction according to the difference between the actual contact state and the simulated contact state. S5. After detecting the joint transmission dead zone, apply a small-amplitude alternating compensation signal to the corresponding joint. S6. After the contact force exceeds the safety boundary, the control end effector retracts in the opposite direction and performs a state reset and degraded retry; This method is applicable to precision contact-based industrial operations in high-end manufacturing production lines. It is used to transfer industrial operation skills learned by embodied robots in a 3D simulation environment to the execution process on a real physical production line. Examples include the assembly of aviation plugs under micro-tolerance conditions, blind insertion of high-voltage wiring harnesses in new energy vehicles, precision interface connection, and constant-force bonding and grinding of complex curved surfaces. The executing entity includes at least an industrial robotic arm with multi-degree-of-freedom motion capabilities, a force sensing acquisition device located at the end effector, a vision acquisition device for acquiring spatial information of the target environment and target workpiece, a body state acquisition device for acquiring robot joint position, joint speed, and motor response information, and an edge computing node communicating with the robot controller. The input information includes at least the baseline control strategy formed by simulation training, images or point cloud data output by the vision acquisition device, and contact force and contact measurement data output by the force sensing device. The method utilizes torque data, joint position, joint velocity, joint current, and end-effector pose data output by the robot's body state acquisition device, as well as spatial position information of the target area corresponding to the task and process safety boundary parameters. Addressing the dynamic differences between 3D simulation and real physical contact, and the asynchrony issues of multimodal data in sampling period, transmission link, and processing delay, the method performs online correction of the simulation prior strategy. This ensures that the robot executes control according to a unified time and spatial reference during target approach, initial contact, continuous contact, and abnormal contact. The output results include at least synchronized multimodal state information, dynamically adjusted perception fusion weights, online corrected contact control parameters, joint control commands after transmission dead zone compensation, and retreat and retry control actions in over-limit scenarios, enabling the execution process in the real production line to form a continuous and complete control closed loop.

[0019] Specifically, such as Figure 2 As shown: After the robot receives a contact-type operation task from the production line scheduling system, it first accesses the baseline control strategy corresponding to the current task and simultaneously establishes a multimodal historical state sequence. The prerequisite for triggering this process is that the robot controller completes task loading and enters the execution state, and the current task identifier, end-effector coordinate parameters, workpiece type identifier, and strategy file in the strategy storage area are consistent. The edge computing node reads the corresponding baseline control strategy according to the task identifier and loads the strategy parameters into the running memory. The baseline control strategy adopts a control strategy formed by offline training for contact-type operations in a 3D simulation environment. During the training phase, domain randomization can be introduced to allow workpiece position deviation, surface friction changes, joint response differences, and contact stiffness changes to participate in training within a predetermined range. After offline verification, it is put into use in the real production line. The reason for adopting this method is that the contact state in the real production line is related to the task type, workpiece type, and end-effector state. Only by accurately matching the task with the strategy can we ensure that the subsequently established historical state sequence and time alignment relationship have a clear control object. Visual data, force data, and robot body state data differ in sampling period, transmission link, and processing latency. Directly combining raw data within the same control cycle can easily lead to inaccurate time correspondence, causing pre-contact position determination, instantaneous contact response, and post-contact parameter adjustment decisions to be based on different time slices. To avoid this problem, the controller first establishes a unified reference time source and then constructs a historical state sequence based on this unified reference time source. The unified reference time source can be selected as the controller's master clock or a fieldbus synchronization clock. When the visual acquisition device, force acquisition device, and body state acquisition device have local timestamp writing capabilities... When the data acquisition is complete, each acquisition device writes the acquisition time. When the acquisition device does not have local time writing capability, the controller writes the arrival time when the data arrives at the receiving end and corrects it according to the pre-calibrated fixed link delay. The calibration method for the fixed link delay can be to perform multiple round-trip sampling tests under no-load conditions, with a continuous test count of no less than 100 times, and then take the average delay as the compensation value. This processing can ensure that when the robot body state is traced back according to the visual acquisition time, the effective time information under the same time reference is used, rather than the disordered data of mixed acquisition time and arrival time. In real industrial settings, after completing exposure, image transmission, depth calculation, and feature extraction, the effective frequency of the visual state information required for control decisions by the vision acquisition device is typically in the range of 30 Hz to 60 Hz. This range is consistent with the common output levels of industrial-grade depth cameras, structured light cameras, and RGB-D cameras after preprocessing on the production line, and can meet the refresh requirements for macroscopic target localization and approach posture guidance. Force sensing devices and robot body state acquisition devices need to reflect millisecond-level changes at the moment of contact, and their sampling frequencies are typically between 500 Hz and 2000 Hz. This range is consistent with the bandwidth of the six-dimensional force sensor, the servo feedback cycle, and the real-time control link of the industrial robot. This configuration aligns with common practices. To balance the state resolution of contact-based operations with the real-time processing load of the controller, the robot's state sampling frequency can be selected as 1000 Hz, corresponding to a 1-millisecond control cycle. The rationale for choosing 1000 Hz is twofold: firstly, a 1-millisecond cycle can cover the state acquisition needs during contact establishment, contact transition, and micro-manipulation adjustment phases; secondly, 1000 Hz is a common real-time update frequency for industrial robot controllers and servo systems, facilitating compatibility with existing industrial hardware platforms. In high-dynamic contact tasks, the body state sampling frequency can be increased to 2000 Hz; in low-speed approach or low-impact tasks, the body state sampling frequency can be reduced to 500 Hz. After establishing a unified time stamping rule, the controller uses a rolling recording method to create a historical state buffer, continuously writing force data and robot body state data into this buffer in chronological order. The historical state buffer can be configured as a circular queue structure to enable sequential writing, overwriting updates, and fast addressing within a fixed storage space. When the body state sampling frequency is 1000 Hz, the length of the historical state buffer can be selected as 2048 sampling points, corresponding to a historical time span of 2.048 seconds. The selection of a buffer length of 2048 sampling points is based on the following: the cumulative latency of visual data from exposure, transmission, preprocessing to output of visual state information typically does not exceed 150 milliseconds under normal operating conditions, while network jitter and buffer overflows can significantly increase the latency. During task switching transients, visual link latency may be amplified briefly. Therefore, setting the historical span to more than 2 seconds allows for a time backtrack margin of more than ten times the normal visual latency, preventing short-term congestion from causing historical state matching failures. Simultaneously, the 2048-bit binary capacity facilitates fast read / write operations in the controller memory using bit-addressing or modulo-length overwriting for the circular queue. Each record written to the buffer includes at least the following fields: time stamp, joint angle, joint angular velocity, joint drive current, end-effector pose, triaxial force, triaxial torque, and current task identifier. Records in the buffer are arranged in ascending order of time stamps, and the time index corresponding to the most recently written position is retained for subsequent retrieval based on the principle of minimum time difference. To ensure the reliability of historical state sequences, the controller continuously monitors the time interval between two adjacent frames of data from each acquisition device and compares the actual time interval with the rated sampling period. If the time interval between two adjacent frames of any acquisition device deviates from the rated sampling period by more than 20% consecutively, the acquisition device is marked as an abnormal data source, and clock resynchronization is triggered. The 20% threshold for anomaly determination has an engineering basis: the periodic jitter of an industrial acquisition link under normal conditions is generally less than 10% of the rated sampling period, and 20% is significantly higher than the normal jitter range, which can effectively distinguish between occasional jitter and persistent problems such as clock drift, link blockage, and acquisition anomalies. During resynchronization, the controller resends the synchronization clock to the abnormal data source and verifies the time interval in subsequent control cycles. If the anomaly persists for three consecutive sampling periods after resynchronization, the control command issuance of the current baseline control strategy is suspended, only the robot standby state is retained, and a time synchronization anomaly flag is output to the upper control system. The basis for using three consecutive sampling periods as the stop condition is that single and double anomalies may still be instantaneous link disturbances, while three consecutive sampling periods of anomalies better reflect synchronization failure or persistent link anomalies, thereby reducing the risk of false and missed stops. After the abnormal data source completes resynchronization, the controller continues to monitor its output time interval. If the data source returns to normal for five consecutive sampling periods, the abnormality flag is removed, and data access for the current task is restored. Setting the recovery judgment condition to five consecutive normal sampling periods is to make the recovery condition stricter than the abnormality identification condition, avoiding the data source that has just completed resynchronization from entering a fluctuating state again in a short period of time, which would cause repeated task switching. After recovery, newly written data continues to enter the historical state buffer, and old data is overwritten and updated according to the circular queue rule. If a time deviation of more than 20% occurs again during the recovery observation period, clock resynchronization is re-executed, and the recovery strategy command is not directly issued. Through this abnormality handling and recovery mechanism, the historical state sequence can maintain clear continuity boundaries and identifiable effective intervals in abnormal scenarios, thereby ensuring a stable data foundation when performing time matching and spatiotemporal alignment based on the historical state sequence in the future. After the above processing, a multimodal historical state sequence organized according to a unified time base is formed, and the correspondence between the current task and the current baseline control strategy is established. This historical state sequence is used to backtrack and match the corresponding robot body state according to the visual acquisition time, and to support subsequent spatiotemporal alignment processing.

[0020] Specifically, such as Figure 3 As shown: After establishing the multimodal historical state sequence, the controller performs spatiotemporal alignment processing on the visual data and the robot body state data. The conditions for triggering this processing are: the visual acquisition device outputs usable observation results of the target area, and the acquisition time label corresponding to the visual data has been obtained. Since the visual data only reaches the controller after exposure, transmission, calculation, and feature extraction, while the robot body state data is continuously updated at a higher frequency, if the time when the visual data arrives at the controller is directly used as the basis for spatial judgment, the visual observation results are likely to lag behind the current robot's true pose. Therefore, the controller uses the acquisition time of the visual data as a backtracking index to determine the robot body state corresponding to the acquisition time in the historical state sequence, and performs forward correction on the visual observation results based on the body motion between the acquisition time and the current control time to generate the synchronization state corresponding to the current control time. The acquisition time label for visual data can be selected as the camera exposure center time, depth frame generation time, or image sampling completion time, with priority given to the original acquisition time label written by the camera's internal synchronization circuit at the data source. After reading the acquisition time label, the controller searches the historical state sequence for the body state record with the smallest time difference from that acquisition time. When two adjacent body state records exist in the historical state sequence before and after that acquisition time, linear interpolation can be performed based on the two records to obtain the body state corresponding to the intermediate time. Then, it is compared with the nearest neighbor matching result, and the result with the smaller time error is taken as the matching result. If the minimum time difference exceeds two body state sampling periods, the visual data frame is marked as insufficiently aligned. Two body state sampling periods are used as the judgment. The basis for the limit is as follows: when the aforementioned body state sampling frequency is 1000 Hz, a single sampling period corresponds to 1 millisecond, and two sampling periods correspond to 2 milliseconds; when the robot's approach speed is on the order of 100 millimeters per second, 2 milliseconds corresponds to a displacement error of about 0.2 millimeters, which is still within the range of common spatial reconstruction residuals after near-range depth vision calibration; the spatial reconstruction residuals after near-range calibration are usually between 0.2 millimeters and 0.5 millimeters, so using 2 milliseconds as the upper limit of time matching sufficiency can ensure the accuracy of time-matching without frequently triggering invalid degradation due to overly strict judgment conditions; after entering the insufficient alignment state, the controller does not directly discard the visual data of that frame, but retains its directional reference function and reduces its initial visual guidance weight in subsequent control; After completing the time-matching, the controller extracts the continuous body state trajectory between the matching time and the current control time, and performs spatial mapping on the visual observation results based on joint angle changes, joint velocity changes, and end-effector posture changes. The spatial mapping relationship is obtained through the robot calibration model, hand-eye calibration matrix, and the installation transformation relationship between the end-effector and the vision acquisition device. When the vision acquisition device is installed at a fixed workstation, a fixed extrinsic parameter matrix is ​​used to map the visual coordinate system to the robot base coordinate system. When the vision acquisition device is installed at the wrist of the robotic arm, wrist posture changes and end-effector compensation are introduced simultaneously to dynamically transform the visual coordinate system. To reduce the spatial deviation caused by robot motion during the time delay, the controller further... Based on the end-effector pose increment between the matching time and the current control time, the target region position is forward-corrected. Specifically, the end-effector poses corresponding to the matching time and the current control time are calculated separately, and the pose difference between the two is calculated and applied to the target region position under visual observation. For point cloud data, rigid body transformation can be directly applied to the target region point set. For the target region contour in a two-dimensional image, the contour features are preferentially restored to three-dimensional feature points by combining the depth map before applying rigid body transformation. When a depth map is not available, three-dimensional feature points can also be restored by combining binocular parallax results or pre-calibrated geometric constraints of the workpiece working surface. Through the above processing, the visual observation result can be mapped from the position corresponding to the original acquisition time to the position corresponding to the current control time. After completing spatial mapping and forward correction, the controller verifies the confidence level of the target region recognition result. If the target region detection confidence level is below 0.6 for three consecutive frames, the current visual state is marked as a weakly confident state. The basis for using 0.6 as the confidence threshold is that this threshold can be determined based on the balance between precision and recall of the target region recognition model on offline validation samples. In most industrial vision recognition models, the false match rate will increase significantly when the confidence level is below 0.6. The use of three consecutive frames as the judgment condition also has a clear basis: the effective output frequency of visual state information is usually in the range of 30 Hz to 60 Hz, and three consecutive frames correspond to a continuous observation time of about 50 milliseconds to 100 milliseconds. This duration can effectively distinguish between instantaneous reflection and short-term occlusion. This approach avoids occasional low-confidence results caused by random noise, while preventing significant lag in control state switching due to excessively long waiting times. Upon entering a weakly confident state, the controller does not completely block visual information but instead limits the visual guidance weight to 30% to 50% of the initial visual guidance weight under normal conditions. The basis for this limitation range is that during offline calibration and bench testing of contact-type operations, when the visual guidance weight is below 30%, the target direction constraint capability is significantly weakened; when the visual guidance weight is above 50%, low-confidence visual information may still have an excessively strong impact on the control output during the near-contact phase. Using a limitation range of 30% to 50% can reduce the interference of low-confidence visual information on near-contact control while preserving the target direction reference capability. The controller also performs an additional judgment on visual data latency. If the cumulative latency of visual data from the completion of acquisition to the current control moment exceeds 150 milliseconds, the visual information of that frame is judged as lagging visual information and is no longer directly used for close-range contact control, but is only retained for macroscopic direction reference. The basis for using 150 milliseconds as the latency threshold is that when the robot's approach speed is on the order of 100 millimeters per second, 150 milliseconds corresponds to a displacement error of about 15 millimeters. 15 millimeters is close to the critical distance range for the robot to transition from the macroscopic approach stage to the microscopic contact stage. If the visual result of that frame continues to be used with high weight, it is easy for the end effector to enter the false contact area based on outdated spatial position information. Therefore, when the cumulative latency exceeds 150 milliseconds, the controller downgrades the credibility of the visual data of that frame to a low credibility level and only uses it to maintain the general direction constraint of the target area, and no longer uses it for close-range contact judgment and small pose correction. After the above processing, the controller generates the synchronization state corresponding to the current control moment. The synchronization state includes at least the aligned target area spatial position, target area contour direction, current robot body state, visual confidence indicator, alignment sufficiency indicator, and delay state indicator. When the visual data is in an inadequate alignment state, a weak confidence state, or a lagging visual state, the corresponding indicator is written into the synchronization state for subsequent control process calls. The synchronization state is used for subsequent multimodal weight adjustment, proximity state judgment, and contact control parameter update.

[0021] Specifically, such as Figure 4 As shown: After obtaining the synchronization state corresponding to the current control moment, the controller dynamically adjusts the fusion weights of visual guidance information and force feedback information based on the proximity between the end effector and the target contact area. The conditions for triggering this process are: the end effector has not yet entered a clear contact state, and the synchronization state already contains the available target area spatial location, the current end effector pose, and the visual confidence indicator. The controller does not use a fixed ratio fusion method, but continuously changes the proportion of different modalities in the control decision based on the real-time proximity. When the robot is away from the target contact area, visual information is more suitable for providing the macroscopic direction and relative position of the target area. As the end effector gradually approaches the target contact area, the visual information is affected by depth error, surface reflection, partial occlusion, calibration residual, and small end effector pose drift, and its near-distance geometric judgment stability gradually decreases. Therefore, it is necessary to increase the proportion of force feedback and small end effector motion states in the control. The controller uses the real-time distance between the end effector and the target contact area as the basis for adjusting the fusion weights. This real-time distance can be directly obtained from the spatial position of the target area and the current end effector pose in the synchronization state. When the target contact area is a hole, slot, or curved surface to be contacted, the shortest distance from the center point of the end effector to the reference surface or reference center of the target contact area is preferentially used as the proximity index. According to different proximity levels, the controller divides the approach process into a macro-proximity interval, a transition interval, and a micro-contact preparation interval. When the real-time distance is greater than 15 mm, the end effector is determined to be in the macro-proximity interval; when the real-time distance is between 3 mm and 15 mm, the end effector is determined to be in the transition interval; when the real-time distance is less than or equal to 3 mm, the end effector is determined to be in the micro-contact preparation interval. 15 mm is used as the boundary value between the macro-proximity interval and the transition interval because the measurement noise of industrial depth vision at typical operating distances is usually taken as 2. From millimeters to 3 millimeters, with an end-effector approach speed of 100 millimeters per second, statistical analysis of approach tests under no-load and rated load conditions shows that the cumulative displacement error caused by control command issuance, servo response, and mechanical braking is typically 8 to 12 millimeters. The sum of these two factors forms a combined error range of approximately 10 to 15 millimeters. Therefore, setting 15 millimeters as the upper boundary effectively covers the combined effects of visual and mechanical response errors. Using 3 millimeters as the boundary between the transition zone and the micro-contact preparation zone is based on the following: the end-effector repeatability error during the near-contact stage is typically 0.5 to 1 millimeter; the transmission chain compensation residual is typically 0.5 to 1 millimeter; and the sum of end-effector elastic deformation and workpiece local tolerances is typically 0.5 to 1.5 millimeters. The combined error falls within the range of approximately 1.5 to 3.5 millimeters. Therefore, setting 3 millimeters as the boundary where the visual dominance essentially ceases is more consistent with the actual error level during the near-contact stage. When the end effector is in the macroscopic approach range, the controller sets visual guidance information as the primary input and force feedback information as the auxiliary input. The visual guidance weight can be selected from 0.8 to 0.98, and the force feedback weight can be selected from 0.02 to 0.2. Preferably, the visual guidance weight is 0.95 and the force feedback weight is 0.05. This set of preferred values ​​is determined through offline calibration and bench testing. During the calibration process, approach time, end effector attitude convergence error, and false collision rate are used as evaluation indicators to compare different weight combinations. When the visual guidance weight is lower than... At a weight of 0.8, the macroscopic pose convergence speed of the end effector towards the target area decreases significantly. When the visual guidance weight is increased to around 0.95, it can maintain a certain proportion of force feedback anomaly detection capability while maintaining high approach efficiency. When the force feedback weight is around 0.05, the force feedback mainly undertakes the anomaly detection function without prematurely interfering with the macroscopic approach trajectory. In this range, the controller mainly performs pose adjustment approach based on the target area position and uses force feedback information as safety monitoring information to identify accidental collisions, abnormal dragging, or unintended contact. When the end effector enters the transition zone, the controller continuously adjusts the visual guidance weight and force feedback weight based on the real-time distance, without using abrupt switching. The weight change rule can be a piecewise linear rule, a lookup table rule, or a smooth function rule, with cosine smoothing being preferred. The reason for using cosine smoothing is that the weight change rate is lower at the ends of the interval and higher in the middle, which can balance switching smoothness and response speed. In specific processing, the proportion of the current real-time distance within the 15 mm to 3 mm range is first calculated, and then the visual guidance weight and force feedback weight are calculated according to the cosine function based on this proportion. Weights; when the real-time distance decreases from 15 mm to 3 mm, the visual guidance weight smoothly decreases from 0.95 to 0.05, while the force feedback weight smoothly increases from 0.05 to 0.95. These endpoint values ​​are chosen as the preferred values ​​for the two ends of the transition interval because: the starting point of the transition interval should maintain the same main visual control mode as the macroscopic approach interval, while the ending point of the transition interval is close to the actual contact establishment position, and visual information only needs to retain a very low proportion of directional constraint; force feedback should basically play a dominant role. Using a continuous change method can reduce posture oscillations and control jumps caused by sudden weight changes near the boundary. When the end effector enters the micro-contact preparation zone but has not yet detected clear contact, the controller continues to maintain a low proportion of visual guidance, but no longer uses visual information as the primary control basis. At this time, the visual guidance weight is no higher than 0.1, preferably between 0.05 and 0.1. Force feedback information and the micro-motion state of the end effector are used as the primary inputs. The reason for limiting the visual guidance weight to no higher than 0.1 is that: within the micro-contact preparation zone, the visual error and the overall end effector error are already at similar levels, and further increasing the proportion of visual guidance will amplify the outdated or distorted geometric guidance effect; at the same time, in the offline contact success rate test during the micro-approach phase, when the visual guidance weight exceeds 0.1, the false collision rate and overcorrection rate increase significantly, while controlling it within the range of 0.05 to 0.1 can still maintain the target direction constraint capability; when in this range, the controller mainly determines whether to enter the actual contact state based on the micro-pose change of the end effector, the pose change rate of the end effector, and the contact force increment. In addition to distance conditions, the controller also introduces force pre-trigger conditions as an auxiliary triggering factor for fusion weight switching. When the absolute value of the normal contact force is higher than 0.5 Newtons, even if the distance estimation result is still slightly greater than 3 mm, the controller will increase the control ratio of force feedback information in advance. The basis for using 0.5 Newtons as the pre-trigger threshold is that this threshold comprehensively considers the zero drift range of the six-dimensional force sensor, the gravity compensation residual of the end tool, and the upper limit of workshop vibration noise. It can avoid misidentifying noise as contact and complete the control center of gravity transfer in time when the initial contact just occurs. In order to avoid the instantaneous disturbance of a single sampling point from mis-triggering this condition, it is preferred to require the normal contact force to be higher than 0.5 Newtons for 3 consecutive control cycles before performing the early switching. When the control cycle value is 1 millisecond, 3 consecutive control cycles correspond to 3 milliseconds. This duration is sufficient to filter out occasional noise and will not significantly delay the control mode switching before contact. When visual guidance information and force feedback information conflict, the controller prioritizes the force feedback information and adds an occlusion or reflection anomaly flag to the visual state. Conflict determination can be performed under the following conditions: the distance estimation result is still greater than 3 mm, and the normal contact force is higher than 0.5 Newtons for three consecutive control cycles; or the visual confidence flag is already in a weak or low confidence state, and the normal contact force increment maintains a positive growth for five consecutive control cycles. The basis for using five consecutive control cycles as an auxiliary determination condition is that when the control cycle value is 1 millisecond, 5 milliseconds can cover the short-term contact establishment process and can be distinguished from single random disturbances and instantaneous mechanical vibrations. In the conflict state, the controller no longer adjusts the approach trajectory according to the vision-dominated method, but prioritizes correcting the approach control based on the force feedback information to avoid penetration, pressing, or accidental contact caused by visual misleading. The anomaly flag added to the visual state is also written into the current control state for subsequent confidence updates and anomaly branch processing. After the above processing, the controller generates fusion weight reassembly and contact preparation status information corresponding to the current control moment. This result includes at least the fields of visual guidance weight, force feedback weight, micro-contact preparation identifier, modal confidence identifier, and modal conflict identifier. When there are insufficient alignment, weak confidence, delayed vision, or visual and force conflict states, the corresponding identifiers are written into the current control state for subsequent contact control parameter adjustment and anomaly handling. This result serves as the subsequent control input for performing contact control parameter updates and contact anomaly branch processing.

[0022] Specifically, such as Figure 5 As shown: After the end effector enters the contact phase, the controller adjusts the contact control parameters in the contact direction online based on the difference between the actual contact state and the simulated contact state. The conditions for triggering this process are: the end effector has made contact with the target contact area, or the initial contact confirmation conditions have been met. The initial contact confirmation adopts a multi-condition joint judgment method, giving priority to the contact force change rate as the main confirmation index, and combining it with the normal contact force in the contact direction and the relative positional relationship between the end effector and the target surface for verification. Specifically, the controller continuously samples the three translational force components in the six-dimensional force data and calculates the contact force change rate between adjacent control cycles. When the contact force change rate of any translational force component exceeds 2.5 Newtons per millisecond for three consecutive control cycles, and the normal contact force in the contact direction is higher than 0.5 Newtons for three consecutive control cycles, or the end effector along the contact... Contact is considered established when the geometric gap between the direction and the target surface is reduced to within 0.5 mm. The basis for using three consecutive control cycles as the confirmation window is that when the control cycle is 1 millisecond, three control cycles correspond to 3 milliseconds, a duration sufficient to filter out instantaneous disturbances at a single sampling point. The basis for using 2.5 Newtons per millisecond as the contact force change rate threshold is that this threshold is determined by comparing the spectra of no-load contactless tests, workshop environmental vibration tests, and standard contact sample tests. Signals below this threshold mainly exhibit low-frequency disturbances, air disturbances, or cable dragging effects, while signals exceeding this threshold and lasting for more than 3 milliseconds typically correspond to the actual contact establishment process. The basis for using 0.5 Newtons and 0.5 mm as the normal force auxiliary threshold and geometric verification threshold, respectively, is that they are higher than the upper limits of pre-contact noise and zero drift, and match the magnitude of the end-point fine-tuning error in the near-contact stage. After contact is established, the controller no longer uses the default contact parameters from the simulation environment, but instead compares the actual contact state with the simulated contact state in real time. The actual contact state includes at least the following data items: real-time six-dimensional contact force, real-time end pose, real-time approach velocity, contact direction, displacement change after contact establishment, and contact force rate of change. The simulated contact state is preferentially obtained through an offline-generated reference contact trajectory table. The reference contact trajectory table is pre-generated based on the results of the baseline control strategy in the simulation environment and indexed according to the target area number, end pose interval, approach velocity interval, and contact depth interval. The table includes at least the reference contact force, reference contact direction, reference contact displacement, and reference velocity. Data items include attenuation amount and reference contact depth. During operation, the controller first matches the corresponding trajectory cluster according to the current target area number, and then interpolates within the adjacent index interval based on the current end attitude, approach speed, and contact depth to obtain the simulated contact state corresponding to the current working condition. The reason for using an offline reference contact trajectory table instead of directly calling the complete simulation model online is that the former is convenient to obtain the reference contact result quickly within a millisecond-level control cycle, while avoiding the real-time burden brought by online simulation solution. The contact direction is preferentially selected as the normal direction of the target contact area. When the normal direction of the target contact area cannot be stably obtained, it can be selected as the projection direction of the current main contact force direction on the predetermined insertion direction or predetermined bonding direction. The controller constructs a difference evaluation result based on the actual contact state and the simulated contact state. The difference evaluation result is formed by a rule-based hierarchical method, without relying on abstract scoring. Instead, it judges the contact force deviation, velocity attenuation deviation, displacement deviation, and jamming state separately, and then outputs the normal state, first-level out-of-tolerance state, or second-level out-of-tolerance state. The contact force deviation is determined by the ratio between the actual contact force and the reference contact force. The velocity attenuation deviation is determined by the deviation ratio of the actual approach speed to the attenuation of the reference speed. The displacement deviation is determined by the difference between the actual displacement of the end along the contact direction and the reference contact displacement. The jamming state is determined by the condition that the displacement growth is insufficient after contact is established and the contact force continues to accumulate. Optionally, when the actual displacement along the contact direction accumulates to less than 0.05 mm in 5 consecutive control cycles, and the normal contact force continues to rise within the same window, the jamming state is determined to be established. The basis for using 0.05 mm as the lower limit of displacement growth is that this value matches the lower limit of the end position resolution and the encoder calculation error. When it is lower than this value, it can be regarded as that the end has basically not made effective progress along the contact direction. When the actual contact force reaches more than 1.5 times the reference contact force and the velocity decay deviation reaches more than 30% for 5 consecutive control cycles, it is judged as a Level 1 out-of-tolerance state. When the actual contact force reaches more than 2 times the reference contact force and the velocity decay deviation reaches more than 60%, or the displacement deviation of the end along the contact direction reaches more than 1 mm and is accompanied by a stuck state for 5 consecutive control cycles, it is judged as a Level 2 out-of-tolerance state. The basis for using 1.5 times and 2 times as the contact force deviation threshold is that: when the contact force is amplified to more than 1.5 times, it usually indicates that the stiffness of the real environment is significantly higher than the simulation expectation, but has not yet entered the significant instability range; when it is amplified to more than 2 times, the contact mismatch has been significantly aggravated, and continuing to maintain the original contact parameters is likely to lead to continuous pressure and oscillation. The basis for using 30% and 60% as the velocity decay deviation threshold is that they correspond to the moderate and significant decrease in propulsion capability after contact is established, respectively. The basis for using 1 mm as the displacement deviation threshold is that this value matches the upper limit of the end-effector comprehensive error in the near-contact stage. Exceeding this level usually means that the contact position deviation or local stuck has significantly affected the propulsion process. The preferred contact control parameters include compliance parameters and response suppression parameters in the contact direction, which can be represented by stiffness and damping parameters respectively in engineering. When the difference evaluation result is in a normal state, the controller maintains the reference contact control parameters in the current contact direction. When the difference evaluation result enters the first-level out-of-tolerance state, the controller reduces the stiffness parameter in the contact direction and increases the damping parameter. When the difference evaluation result enters the second-level out-of-tolerance state, the controller further reduces the stiffness parameter in the contact direction and continues to increase the damping parameter to enhance compliance and energy dissipation capabilities. The reference stiffness parameter in the contact direction can be selected as 100% of the calibration value in the free space stage, adjusted to 40% to 60% in the first-level out-of-tolerance state, and adjusted to 10% to 20% in the second-level out-of-tolerance state; damping... The parameters were increased from the basic damping value in the normal contact stage to 1.2 to 2 times. The basis for this ratio is as follows: In the first-level out-of-tolerance state, contact mismatch has already formed but can still be absorbed by moderate compliance adjustment. Therefore, reducing the stiffness to 40% to 60% can retain the necessary directional support capacity. In the second-level out-of-tolerance state, the contact mismatch has significantly expanded, and it is necessary to quickly release the pressure tendency. Therefore, the stiffness is further reduced to 10% to 20% to significantly improve compliance. The basis for increasing the damping to 1.2 to 2 times comes from contact response bench tests: when the damping is increased to below 1.2 times, the low-frequency oscillation suppression effect is not obvious; when the damping is increased to more than 2 times, the end-propulsion response slows down significantly and is prone to contact stagnation. Therefore, this range is taken as the preferred adjustment range. In terms of specific engineering quantities, the reference stiffness parameter in the contact direction during the free-space stage can be calibrated to 5000 N / m, and then adjusted to the range of 200 N / m to 500 N / m under the second-level out-of-tolerance condition. The basis for using 5000 N / m as the reference stiffness parameter is that this value was obtained from free-space tracking calibration. When the end-effector mass is between 2 kg and 5 kg and the free-space approach speed is on the order of 100 mm / s, it can stably control the no-load position error along the contact direction within the range of 0.2 mm to 0.5 mm, which is consistent with the commonly used Decathlon equations for small-to-medium load industrial robotic arms in precision approach processes. The stiffness parameters are matched in the direction; the reason for lowering the stiffness parameters under the second-level out-of-tolerance state to the range of 200 N / m to 500 N / m is that when an abnormal contact impact of 10 N occurs, this stiffness range can provide approximately 20 mm to 50 mm of compliant displacement space, thereby avoiding the direct conversion of transient impact into continuous pressure; this compliant displacement range does not exceed the safe yield stroke allowed by the current task, which can prevent the robotic arm from significantly deviating from the target area while maintaining operational safety; by adopting the above-mentioned combined stiffness and damping adjustment method, the contact process can be transformed from a rigid pressure response to a compliant control and damped dissipation response; During the online adjustment of contact control parameters, the controller also monitors the contact oscillation state. The controller first performs a first-order low-pass filter on the contact force signal in the contact direction, with a filter cutoff frequency of 80 Hz to reduce the impact of high-frequency measurement noise on oscillation judgment. Then, using 20 consecutive control cycles as a sliding observation window, the controller calculates the average and peak-to-peak values ​​of the filtered contact force within this window and counts the number of sign changes in the first-order difference of the contact force. If, within 20 consecutive control cycles, the peak-to-peak value of the contact force in the contact direction exceeds 20% of the average contact force in the same window, and the number of sign changes in the first-order difference of the contact force is less than 20%, the controller will take action. If there are fewer than 6 occurrences, the system is considered to be in a potential flutter state. The basis for using 20 consecutive control cycles as the judgment window is that when the control cycle is 1 millisecond, 20 milliseconds is sufficient to cover a local oscillation observation window after contact is established. The combination of peak value exceeding 20% ​​of the average contact force and the number of sign changes being no less than 6 is to distinguish between small fluctuations and continuous high-frequency oscillations in normal contact adjustment. After entering the potential flutter state, the controller further reduces the advance speed along the contact direction and widens the update cycle of attitude correction commands to 5 milliseconds to reduce the coupling amplification effect between attitude correction and contact oscillation. If stable contact cannot be established during the contact phase, the controller outputs a retreat request flag. Stable contact establishment failure can be determined based on the following conditions: the contact force along the contact direction is less than 1 Newton for 50 consecutive control cycles, the target contact depth has not been reached, and the displacement fluctuation amplitude along the contact direction at the end exceeds 0.5 mm. The 50 consecutive control cycles are used as the judgment window because when the control cycle is 1 millisecond, 50 milliseconds can cover the short-term stable observation process after a near-contact advance. The 1 Newton lower limit is used because this value is higher than the pre-trigger threshold of 0.5 Newtons before contact, distinguishing it from the slight touch state before contact. The 0.5 mm displacement fluctuation amplitude threshold is used because this magnitude is close to the upper limit of the allowable error for fine-tuning at the end of the contact phase; exceeding this amplitude usually indicates a deviation in the workpiece position or target direction. The target contact depth can be predetermined based on the geometric model of the target area, task trajectory parameters, or assembly process parameters. After outputting the retreat request flag, subsequent control processes can perform retreat processing and re-approach processing based on this flag. After the above processing, the controller generates the contact control parameter group, difference evaluation result, potential flutter status identifier, and backoff request identifier in the current contact direction. The result includes at least the fields of contact direction stiffness parameter, contact direction damping parameter, difference classification identifier, flutter risk identifier, and backoff request identifier. When there is a first-level out-of-tolerance state, a second-level out-of-tolerance state, or a potential flutter state, the corresponding identifier is written into the current control state for subsequent transmission dead zone compensation judgment and contact anomaly handling.

[0023] Specifically, such as Figure 6 As shown: After the contact control parameters enter the online adjustment state, the controller identifies the joint transmission dead zone and applies a micro-amplitude alternating compensation signal to the corresponding joint when the identification is successful. The conditions for triggering this process are: the robot is in a continuous contact state or a small displacement adjustment state, and during the local fine adjustment process, there are phenomena such as discontinuous end-effector micro-displacement following, continuous increase in joint drive current, and insufficient external position response. The reason for this phenomenon is that the reducer, coupling, and joint transmission chain of real industrial robotic arms generally have backlash, static friction dead zone, and slight stick-slip effect during the reverse micro-motion or low-speed crawling stages. These non-ideal factors are usually simplified in the simulation environment, resulting in the control strategy assuming a smooth and continuous relationship between joint output commands, actual joint displacement, and end-effector spatial displacement. To correct this difference, the controller monitors the difference between the input response and output response of key joints in real time during the continuous contact stage, and applies a micro-amplitude alternating compensation signal to the corresponding joint after determining that the transmission dead zone is successful. The controller prioritizes simultaneously acquiring encoder data from both the joint motor side and the reducer output side, and combines this with the reduction ratio to calculate the equivalent input angular displacement and equivalent output angular displacement, then calculates the difference between the two. If the device does not have an output side encoder, the equivalent input-output difference is constructed using the motor side encoder data and the end effector micro-displacement calculation results. The end effector micro-displacement calculation results are preferentially projected onto the corresponding joint output side using the end effector pose change and the joint Jacobian relationship. The controller also simultaneously acquires joint drive current, joint command direction, and end effector micro-displacement data along the task direction. To identify static friction lockout boundaries, the controller pre-calibrates a static friction identification threshold. The static friction identification threshold can be selected as the threshold when the joint is crawling at low speed under no-load conditions. The driving current is set at 1.1 to 1.3 times the average value of the first continuous output displacement during the trial, with a preferred value of 1.2 times. Continuous output displacement can be defined as: the output side angular displacement increases in the same direction for 5 consecutive control cycles, and the cumulative output angular displacement increment reaches more than 0.01 degrees. The basis for using 0.01 degrees as the lower limit of continuous output displacement is that this value is higher than the output side encoder quantization error and the upper limit of static jitter, while it is still in the initial stage of joint micro-movement. The static friction identification threshold is set at 1.2 times the average value of the first continuous output displacement driving current, because this value is higher than the upper limit of short-term current fluctuation, but lower than the driving current range that causes a large displacement change at the end, making it suitable as an identification condition for static friction lock-up boundary. The controller determines that a transmission dead zone exists when the following conditions are met simultaneously within multiple consecutive control cycles: the joint drive current is higher than the static friction identification threshold, the joint command direction reverses within a small range, the equivalent angular displacement difference is greater than 0.02 degrees, the output side angular velocity is lower than 0.001 radians per second, and the duration of the above states reaches 5 milliseconds. The basis for using 0.02 degrees as the dead zone entry threshold is that this value is determined by superimposing the factory backlash test results of the reducer with the assembly error of the whole machine, which can cover the typical backlash level of a precision reducer near the rated load. The basis for using 0.001 radians per second as the output side angular velocity threshold is that this value is higher than the upper limit of the pseudo-velocity formed by static noise, but lower than the lower limit of the effective micro-motion output velocity, which can distinguish between the pseudo-static state and the real micro-motion state. The basis for using 5 milliseconds as the minimum identification window is that this time length is determined by the current establishment time at the moment of joint reverse switching and the statistical results of the reducer micro-motion response delay. When it is shorter than this window, it is easy to misjudge normal short-term inertial lag as continuous dead zone lock-up. After determining that a transmission dead zone has been established, the controller does not change the current task direction. Instead, it superimposes a micro-amplitude alternating compensation signal onto the basic control command of the corresponding joint, so that the joint transmission chain can release the static friction lock-up state without causing a large-amplitude displacement change at the end. The micro-amplitude alternating compensation signal can be a symmetrically alternating small torque signal or a small-amplitude swing position command, preferably a sinusoidal alternating torque signal. The frequency of the compensation signal can be selected from 30 Hz to 60 Hz, preferably 45 Hz. The amplitude of the compensation signal can be selected from 1% to 5% of the rated continuous torque of the corresponding joint, preferably 2%. The reason for choosing 45 Hz as the preferred frequency is that this value... Based on the results of the modal tests of the robotic arm structure, the frequency range is determined to be below the lower limit of the first-order resonant frequency of the main structure and above the low-frequency servo adjustment bandwidth. This allows for sufficiently fast alternating excitation within the joints and reducers without inducing resonance in the entire machine. The optimal amplitude value of 2% of the rated continuous torque is based on the fact that this value was obtained from the joint static friction identification test. It is higher than the minimum disturbance torque required to overcome the peak static friction and lower than the upper limit of the disturbance torque that would cause visible end-effector jitter or significant displacement abrupt change. For scenarios with different equipment structures, reducer hysteresis, or static friction thresholds, adjustments can also be made within the above frequency and amplitude range based on the results of the modal tests and static friction identification. After the compensation signal is applied, the controller continuously monitors the equivalent angular displacement difference of the target joint, the output-side angular velocity, the continuity of the micro-displacement of the end effector along the task direction, and the change in contact force. If the equivalent angular displacement difference drops below 0.01 degrees, the output-side angular velocity recovers to above 0.003 radians per second, and the micro-displacement of the end effector along the task direction continues to increase in the same direction for five consecutive control cycles, then the transmission dead zone is determined to be released. The basis for using 0.01 degrees as the dead zone release threshold is that this value is half of the 0.02-degree entry threshold, which can form a hysteresis interval between the entry threshold and the release threshold, thereby avoiding frequent switching between dead zone identification and release identification near the boundary. The basis for using 0.003 radians per second as the recovery threshold is that this value is higher than the static... The pseudo-speed level corresponding to the noise reduction is still lower than the normal micro-motion speed in the fine-tuning stage, making it suitable as a criterion for judging whether effective output has been restored. After the dead zone is cleared, the controller gradually withdraws the compensation signal according to the decay window of 10 to 20 milliseconds. The basis for using 10 to 20 milliseconds as the withdrawal window is that this time window is determined by the statistical results of the end displacement stability in the compensation signal withdrawal test. If it is shorter than 10 milliseconds, the compensation withdrawal is too fast and it is easy to cause secondary disturbances. If it is longer than 20 milliseconds, it will slow down the normal fine-tuning response. To ensure the safety of the compensation process, if the contact force in the contact direction exceeds the upper limit of the current task's allowable contact force during the compensation period, or if the joint drive current reaches the driver protection threshold, the compensation will be terminated immediately and a failure flag will be output. If the duration of the compensation signal application exceeds one complete excitation cycle, and the equivalent angular displacement difference is still greater than 0.02 degrees, the output side angular velocity is still less than 0.001 radians per second, and the average value of the joint drive current within the observation window is consistently higher than the static friction identification threshold, and the average current value of the last 5 control cycles of the window increases by more than 10% compared to the average current value of the first 5 control cycles of the window, then the controller determines that the dead zone compensation has failed. For the preferred frequency of 45 Hz, one complete excitation cycle corresponds to approximately 22 milliseconds, so the compensation failure observation window can be selected as 20 to 25 milliseconds. The basis for setting this window value to at least one complete excitation cycle is that it is necessary to ensure that the alternating compensation signal has fully exerted its effect on the static friction lock in both the positive and negative directions. If the effective output is not restored after the end of the window, it indicates that the current abnormality is not simply a transmission dead zone, but is more likely related to workpiece hard jamming, contact direction deviation, or local mechanism interference. The basis for using an average current increase of more than 10% as an auxiliary judgment condition is that this proportion is higher than the normal current fluctuation range, which can reflect that the resistive load is still continuously increasing during the compensation application period. After the above processing, the controller generates the current joint's compensated state information; the result includes at least the target joint identifier, equivalent angular displacement difference, output side angular velocity, compensation signal frequency, compensation signal amplitude, dead zone release identifier, and dead zone compensation failure identifier; when it is detected that the transmission dead zone has been released or the dead zone compensation has failed, the corresponding identifier is written into the current control state for subsequent contact anomaly handling and retry judgment.

[0024] Specifically, such as Figure 7 As shown: After detecting that the contact force exceeds the process safety boundary, the controller performs reverse yielding, state reset, and downgraded retry processing. The conditions for triggering this process are: the absolute value of the contact force in any control cycle exceeds the process safety boundary corresponding to the current task; or a dead zone compensation failure flag has been output and the contact force continues to rise; or a continuous chatter risk flag has been formed and has not been eliminated after local parameter adjustment. The process safety boundary is related to the material strength of the workpiece, the load-bearing capacity of the clamping structure, and the allowable impact level of the production line. For precision connector scenarios, the total contact force safety boundary can be selected as 35 Newtons. The basis for using 35 Newtons as the safety boundary is that this value is obtained through destructive testing statistics for the material and connector structure of the corresponding precision connector, corresponding to the safe range before the structure enters irreversible damage. The actual control upper limit is preferably taken as 70% to 80% of the irreversible damage threshold, which can still retain sufficient safety margin in the presence of normal error fluctuations and transient impacts. For workpieces with different materials, sizes, or connector structures, the process safety boundary can be recalibrated based on the corresponding destructive test results. When the contact force exceeds the process safety boundary, the controller immediately cuts off the continuous output of the current position advance command and initiates a reverse retreat action. The retreat direction is not based on a fixed world coordinate axis direction, but is determined according to the current contact force vector or the current contact normal estimation result, preferentially taking the opposite direction of the current contact force vector as the retreat direction to ensure that the retreat path is opposite to the dangerous contact direction. During the retreat process, the controller adjusts the contact control parameters in the contact direction to the minimum rigidity holding state, retaining only gravity compensation and necessary attitude stabilization control, so that the end effector has compliance capability when leaving the dangerous contact area. The retreat speed can be selected as 15 mm / s, and the maximum retreat distance can be selected as 50 mm. The reason for choosing 15 mm / s as the retreat speed is that this speed is higher than the typical fine-tuning speed in the small displacement adjustment stage, while being lower than the minimum safe distance in the workspace, end effector inertia, and control. The upper limit of secondary collision risk is jointly determined by the cycle; through comparison of retreat speed group tests, it is found that at this speed, it can quickly leave the dangerous contact zone without impacting adjacent structures due to excessive retreat; the basis for using 50 mm as the maximum retreat distance is that this distance is greater than the sum of the maximum characteristic dimensions of the contact establishment zone, the accidental contact zone, and the local interference zone in the current task, while being less than the safe retreat stroke allowed by the workspace. After task trajectory analysis and tooling space measurement, it can provide a new safe starting position for rescanning; if the contact force drops below 0.5 Newtons during the retreat process, early termination of retreat is allowed; the basis for using 0.5 Newtons as the early termination threshold is that this value is consistent with the contact pre-trigger threshold and can serve as a unified judgment boundary for basic contact release; if the retreat distance reaches 50 mm but the dangerous contact zone is not completely left, a strong abnormal status indicator is output, and automatic retry is suspended; After the retreat is completed, the controller performs a state reset process instead of directly restoring the original approach trajectory. The state reset includes at least the following: clearing the contact data segments before the limit exceedance in the historical state buffer, re-establishing the spatial relationship between the current workpiece position and the end position, re-triggering the target area scan of the vision acquisition device, and restoring the contact control parameters, fusion weights, and compensation signal states to the safe initial settings. To improve the pertinence of subsequent retry parameter adjustments, the controller also performs a retrospective analysis of the operating state before the limit exceedance occurs, and records the synchronization state, contact difference evaluation results, fusion weight change process, joint dead zone identification results, and contact force direction at the moment of retreat triggering within 500 milliseconds before the limit exceedance as abnormal samples. The basis for using 500 milliseconds as the retrospective window is that, under the typical approach speed conditions of the current task, this time length can cover the complete local evolution process from macroscopic approach, entering the transition control zone to dangerous contact, which is convenient for identifying the causes of instability caused by visual lag, contact mismatch, dead zone locking, or local interference. After the state reset is completed, the controller decides whether to allow automatic retry based on the rescan results. If the visual rescan shows that the target area offset does not exceed 3 mm, and the number of automatic retry attempts for the current task has not reached the upper limit, a degraded retry is initiated. The basis for using 3 mm as the allowable offset is that this value is consistent with the boundary value of the near-contact preparation interval, and is higher than the visual repositioning error but lower than the upper limit of the allowable geometric deviation in the near-contact stage. After the visual repositioning error statistics and workpiece repositioning deviation statistics are completed, 3 mm is still within the controllable approach range. The degraded retry does not directly repeat the original parameters, but instead calls the conservative retry parameter group calibrated by the combined over-limit working condition playback and group retry test, and synchronously adjusts the approach speed and mode switching edge. The approach speed can be reduced to 50% of the original set value, and the upper boundary of the switch from vision-guided to force-driven and the softening trigger boundary of the contact control parameters can be advanced by 3 mm, i.e., from 15 mm to 18 mm. The reason for reducing the approach speed to 50% is that this ratio can significantly reduce the impact energy during re-contact without significantly increasing the retry time. The reason for advancing the above two boundaries by 3 mm is that this advance comes from the sum of the upper boundary of the visual measurement error and the additional safety margin in the near contact stage, so that the controller can switch to the force-driven state earlier and enter the compliant adjustment state earlier, thereby reducing the risk of re-exceeding the limit caused by visual errors, workpiece deviations or local interference. After a degraded retry is completed, if the contact force does not exceed the process safety boundary again, subsequent operations are allowed to continue. If the limit is triggered again after three consecutive degraded retryes in the same target area, the current workpiece or current station is marked as being in a manual intervention state, and automatic execution is stopped. The basis for using three consecutive degraded retryes as the manual intervention threshold is that one or two failures may still be caused by occasional disturbances, short-term occlusion, or instantaneous attitude deviations, while three consecutive limit violations in the same area can usually rule out occasional factors and be attributed to workpiece size deviations, assembly posture errors, or environmental anomalies. Therefore, it is suitable as a judgment threshold for automatic control to exit and switch to manual intervention. After entering the manual intervention state, the controller outputs a process abnormality code to the upper-level system and retains the corresponding abnormality sample for subsequent investigation and parameter revision. After the above processing, the controller generates an abnormal closed-loop processing result; the result includes at least the following fields: yield completion flag, state reset completion flag, whether retry is allowed flag, conservative retry parameter group, strong abnormal state flag, and manual intervention processing state flag; if the automatic retry is successful, the subsequent control process continues to execute according to the reset synchronization state and conservative retry parameter group; if the automatic retry fails, the current automatic operation cycle is terminated and the manual intervention processing state is maintained.

[0025] Example 2: Based on Example 1, the specific application process of a method for transferring and training industrial operation skills for embodied robots is further explained: Specifically, such as Figure 8 The following description uses the blind insertion task of high-voltage wiring harnesses in the final assembly line of new energy vehicles as an example. The objects to be executed are a high-voltage wiring harness plug with multi-core metal terminals and a corresponding socket installed on the battery pack side. There are small assembly tolerances between the plug and the socket, the visible space in the insertion area is limited, and there are easily interfered structures such as sheaths, brackets, and adjacent wiring harnesses around the socket. In this scenario, a flexible gripper is installed at the end of an industrial robotic arm, a six-dimensional force sensor is set between the gripper and the robotic arm flange, a near-field depth vision camera is installed on the wrist of the robotic arm, a global vision acquisition device is set on the outside of the workstation, and the robot body controller is connected to the edge computing node. Before the operation starts, the edge computing node retrieves the baseline control strategy corresponding to the blind insertion task of the high-voltage wiring harness from the strategy storage area, and completes strategy matching according to the current task number, end tool parameters, and plug model. This baseline control strategy is a control strategy pre-trained by the robot in a three-dimensional simulation environment around the blind insertion action of the high-voltage wiring harness, which is used to provide the basic control framework for free space approach, near-contact fine-tuning, and post-contact propulsion. After the task is initiated, the controller first establishes a unified time reference and incorporates visual data, force data, and robot body state data into a unified timeline for management. Specifically, the controller uses the master clock as a unified reference time source and timestamps the data output from the external visual acquisition device, wrist visual acquisition device, force acquisition device, and joint state acquisition device. For high-frequency force data and joint state data, the controller writes them into the historical state buffer in chronological order. For visual data, the source acquisition time is retained for subsequent execution time backtracking and matching. Since in the blind-plug scenario, external vision primarily provides the macroscopic position of the socket area, wrist vision primarily provides the local alignment of the plug and socket, while force and body state provide rapid feedback at the moment of contact, multimodal data naturally differ in sampling frequency and link latency. By establishing a unified reference time source and maintaining the historical state sequence, the timing inaccuracies caused by directly splicing asynchronous data can be avoided, ensuring that all subsequent control decisions are based on a unified time reference. After the wrist vision or external vision outputs a valid observation result for the current socket area, the controller extracts the acquisition time of the corresponding visual frame and searches the historical state buffer for the robot body state record with the smallest time difference from that acquisition time. If the time difference is within the allowable range, the historical body state is used as the matching state for the visual observation. If the time difference exceeds the allowable upper limit, the proportion of the visual information of that frame in subsequent control is reduced. After completing the time matching, the controller further performs spatial forward correction on the visual observation results based on the body pose change between the matching time and the current control time. For example, when the high-voltage harness plug has moved forward along the insertion direction... While the visual image remains at an earlier time, the controller performs a forward mapping of the socket center position and insertion direction based on the end effector pose increment, ensuring that the current visual result corresponds to the robot's actual pose. For key local features such as socket edges, guide ramps, and foolproof keys, the controller prioritizes 3D reconstruction using the wrist depth map. When the depth map is partially missing, it is supplemented by combining external vision and workpiece geometric constraints, thus forming a synchronized state at the current control moment. This synchronized state includes at least the socket center position, insertion direction, current end effector pose, visual confidence indicator, and alignment sufficiency indicator, which are used for subsequent control logic calls. After synchronization is established, the robotic arm drives the wire harness plug to gradually approach the socket. At this point, the controller does not use a fixed ratio to fuse visual guidance information and force feedback information. Instead, it dynamically adjusts the proportion of visual and force information in the control based on the real-time distance between the plug tip and the target contact area of ​​the socket. When the plug is far from the socket, the controller maintains visual dominance to quickly complete the center alignment of the socket and the correction of the insertion direction. As the plug gradually approaches the socket, the controller gradually reduces the weight of visual guidance and increases the proportion of force feedback and end-effector micro-pose change information, making the system rely more on high-frequency body and force information for fine-tuning during the close contact phase. This dynamic switching is particularly important in wire harness blind insertion scenarios because the socket's surrounding sheath can easily cause partial obstruction, and the metal terminals and plastic shell surfaces can also cause reflection and depth distortion. If visual information is continuously relied upon with high weight during the close contact phase, it can easily cause the plug tip to enter at an angle, press against the socket edge, or accidentally touch the terminals. By continuously switching the fusion weight according to distance, the controller can maintain the advantage of visual guidance during the macro positioning phase and switch to force dominance in advance during the close contact phase, thereby improving the alignment stability in blind insertion scenarios. After the plug tip makes initial contact with the socket inlet, the controller no longer uses the default contact parameters from the simulation environment. Instead, it compares the actual contact state with the simulated contact state in real time. Specifically, the controller determines whether the current real-world working condition is harder or softer than the simulation expectation, or whether local misalignment or jamming has occurred, based on the actual contact force, velocity decay, contact displacement, and simulation contact information in the reference contact trajectory table in the current insertion direction. If the actual contact force is significantly higher than the reference contact force, and the propulsion velocity decays significantly faster than the simulation expectation, it indicates that the actual stiffness between the plug and socket is higher, or that local misalignment has caused the plug tip edge to rub against the socket inlet. At this point, the controller reduces the stiffness parameter in the insertion direction and increases the damping parameter, making the plug more compliant in the contact direction. If the actual contact force is lower than the simulation expectation but the advancement depth is insufficient, it indicates that the current state may be shallow contact, slip contact, or the guide is not properly established. The controller maintains a low advancement speed and retains a certain directional support capability to continue searching for a stable guide relationship. Through this online adjustment method based on the difference between the actual contact state and the simulated contact state, the controller no longer directly applies the simulated contact model to the real production line, but enables the contact control parameters to adapt to the current actual contact state, thereby reducing the risk of plug front-end pressure, terminal deformation, and socket damage. During blind insertion of wire harnesses, when the plug has entered the socket but the pushing action becomes discontinuous, or when the robotic arm has issued a small pushing command but the plug tip does not show a corresponding continuous micro-displacement, the controller further monitors whether the joint drive chain has entered a dead zone state. Since high-voltage wire harness plug insertion is often accompanied by reverse fine-tuning and low-speed, small pushing, the backlash, static friction, and micro-stick-slip effects in the joint reducer and drive chain are easily amplified at this stage, causing the robotic arm current to rise while the end effector remains stationary, intermittent end effector micro-movements, or attitude fine-tuning failure. To address this, the controller prioritizes comparing the difference between the motor-side angular displacement and the output-side angular displacement, and combines this with the joint drive current, the continuity of the end effector micro-displacement, and the reverse switching of commands to determine whether a transmission dead zone has formed. Once the dead zone is confirmed, the controller, without changing the macroscopic insertion direction, superimposes a micro-amplitude alternating compensation signal on the corresponding joint, so that the joint can release the static friction lock-up state without causing significant end-effector advance. During the compensation process, the controller continuously monitors whether the joint output response recovers, whether the end-effector micro-displacement becomes continuous again, and whether the contact force remains within a safe range. If the dead zone is released, the compensation signal is gradually withdrawn. If there is still no effective output recovery after a complete excitation cycle of compensation, and the current continues to increase, it is considered that the current anomaly is more likely to be caused by plug hard jamming, insertion direction deviation, or structural interference, and no longer belongs to a simple joint transmission dead zone. At this time, the controller outputs a dead zone compensation failure flag to provide a basis for subsequent anomaly retraction processing. If, during blind insertion of the wire harness, the contact force exceeds the process safety boundary, the contact force continues to rise after dead zone compensation fails, or contact chatter persists and cannot be eliminated through local parameter adjustment, the controller triggers abnormal retraction. For high-voltage wire harness plug scenarios, the process safety boundary can be determined based on destructive testing of the plug housing, metal terminals, and socket guide structure, and a safety margin is maintained during normal control. When abnormal retraction is triggered, the controller immediately stops further advancement and performs reverse retraction according to the opposite direction of the current contact force vector. During the retraction process, only gravity compensation and necessary attitude stabilization control are maintained to prevent the robotic arm from re-entering the dangerous contact area. Apply additional pressure; if the contact force drops below the contact release threshold, early termination of the retreat is allowed; if the retreat reaches the maximum safe distance but the danger zone is not cleared, a strong abnormal status indicator is output and automatic retry is suspended; after the retreat is completed, the controller does not directly restore the original path, but performs a state reset, clears the contact data fragments before the limit was exceeded, re-triggers the target area scan, restores the contact control parameters, fusion weights and compensation status to the safe initial settings, and records the synchronization status, contact difference evaluation results, fusion weight change process, dead zone identification results and contact force direction at the retreat trigger time during the period before the limit was exceeded as abnormal samples for subsequent cause analysis; After resetting the state, if the rescan results indicate that the offset of the target area of ​​the socket is still within a controllable range and the number of automatic retries has not reached the upper limit, the controller initiates a degraded retry. The degraded retry no longer uses the original proximity parameters, but instead calls the conservative retry parameter group, reduces the proximity speed, advances the visual exit and force-dominated switching boundaries, and simultaneously advances the contact parameter softening trigger boundary, allowing the robotic arm to enter a conservative proximity and compliant contact mode at an earlier stage. In the scenario of blind insertion of wire harnesses, this means that when the robotic arm makes its second or third attempt, it no longer pursues the original efficiency, but prioritizes avoiding impact exceeding the limit again. If the contact force does not exceed the process safety boundary again after the degraded retry, the subsequent insertion action continues. If multiple degraded retry attempts near the same insertion area still exceed the limit again, the controller marks the current workpiece or station as a manual intervention state and outputs a process abnormality code to the upper-level system. In this way, a complete operating closed loop can be formed for deployment on a real production line.

[0026] It should be noted that this invention can be deployed on the device itself to realize embedded applications, or it can run on a PC or other terminal with a user interface, thereby meeting various hardware environments and usage requirements.

[0027] The above embodiments can be implemented in whole or in part by software, hardware, firmware, or any other combination. When implemented in software, the above embodiments can be implemented in whole or in part by a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the processes or functions of the embodiments of this application are implemented in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted wirelessly or wiredly from one website, computer, server, or data center to another website, computer, server, or data center. Wired methods include optical fiber, twisted pair, coaxial cable, etc. Wireless methods include infrared, microwave, etc. Available media include any available media that can be accessed by a computer or data storage devices such as servers and data centers that contain one or more sets of available media. Available media can be magnetic media (floppy disks, hard disks, magnetic tapes), optical media (DVDs), or semiconductor media. Semiconductor media can be solid-state drives.

[0028] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0029] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for transferring industrial operation skills to embodied robots, characterized in that, include: S1. Obtain the baseline control strategy formed by simulation training, and uniformly time-stamp the visual data, force data and robot body state data to establish a historical state sequence. S2. Based on the acquisition time of the visual data, determine the corresponding robot body state from the historical state sequence, and perform spatiotemporal alignment processing on the visual data. S3. Adjust the fusion weight of visual guidance information and force feedback information according to the proximity between the end effector and the target contact area; S4. After detecting that the end effector has made contact with the target contact area, adjust the contact control parameters of the contact direction according to the difference between the actual contact state and the simulated contact state. S5. After detecting the joint transmission dead zone, apply a small-amplitude alternating compensation signal to the corresponding joint. S6. After the contact force exceeds the safety boundary, the control end effector retracts in the opposite direction and performs a state reset and downgrade retry.

2. The method for transferring industrial operation skills to embodied robots according to claim 1, characterized in that, S1 includes: Complete the strategy file matching and load the baseline control strategy according to the task identifier, end tool coordinate parameters and workpiece type identifier; Using the controller master clock or fieldbus synchronous clock as a unified reference time source, the acquisition time is written to the acquisition data that has the ability to write local timestamps, and the arrival time is written to the acquisition data that does not have the ability to write local timestamps, and the link fixed delay correction is performed. Force data and robot body state data are recorded in chronological order using a circular buffer. Based on the comparison between the time interval between adjacent frames and the rated sampling period, abnormal data source identification, clock resynchronization, continuous monitoring, and access recovery are performed.

3. The method for transferring industrial operation skills to embodied robots according to claim 1, characterized in that, S2 includes: Read the acquisition time label of the visual data, and use nearest neighbor matching and adjacent state interpolation in the historical state sequence to determine the corresponding robot body state; By combining the robot calibration model, hand-eye calibration matrix and visual acquisition device installation transformation relationship, spatial mapping and pose advance are performed on the point cloud data or depth-restored image contour data of the target area; The synchronization state is formed by marking the time matching status, the confidence level of target area recognition, and the cumulative delay of visual data.

4. The method for transferring industrial operation skills to embodied robots according to claim 1, characterized in that, S3 includes: Determine the real-time approach distance based on the spatial location of the target contact area and the current end position; When the target contact area is a hole, groove or curved surface, the shortest distance from the center point of the end tool to the reference surface or reference center is used as the approach degree index, and the approach process is divided into macro approach interval, transition interval and micro contact preparation interval. The visual guidance weight and force feedback weight are continuously adjusted according to the smoothing rule; In the micro-contact preparation zone, visual guidance weight is reduced while the proportion of force feedback and end-effector micro-motion control is increased. Modal state marking is performed by combining force pre-triggering conditions and visual-force conflict determination to generate fusion weight reorganization and contact preparation state information.

5. A method for transferring industrial operation skills to embodied robots according to claim 4, characterized in that, In the micro-contact preparation zone, visual guidance is given less weight while force feedback and control of subtle end-effector movements are increased, including: When the end effector enters the micro-contact preparation zone but has not yet detected clear contact, the visual guidance weight is limited to a preset low value range to preserve the directional guidance effect of visual information. Force feedback information and the micro-motion state of the end effector are used as the main control basis, and the contact preparation state is determined based on the micro-positional change of the end effector, the rate of change of the end effector position, and the increase of the contact force.

6. The method for transferring industrial operation skills to embodied robots according to claim 1, characterized in that, S4 includes: Contact confirmation is achieved by combining the rate of change of contact force, normal contact force, and the relative positional relationship between the end and the target surface; Match the reference contact trajectory according to the target area number, end attitude, approach velocity, and contact depth; Differences are classified based on contact force deviation, velocity attenuation deviation, displacement deviation, and jamming state; Adjust the stiffness and damping parameters in the contact direction according to the classification results; The system monitors and determines the contact oscillation state and the stable contact establishment state, and outputs a flutter risk indicator or a backoff request indicator.

7. A method for transferring industrial operation skills to embodied robots according to claim 6, characterized in that, Differences are classified based on contact force deviation, velocity attenuation deviation, displacement deviation, and jamming state, including: The difference evaluation results are constructed using a rule-based hierarchical approach; The ratio of the actual contact force to the reference contact force is used to characterize the contact force deviation. The ratio of the actual approach speed to the reference speed attenuation is used to characterize the speed attenuation deviation. The difference between the actual displacement of the end along the contact direction and the reference contact displacement is used to characterize the displacement deviation; Insufficient displacement growth after contact is established and continuous accumulation of contact force are used to characterize the stuck state. Based on the above criteria, the conditions are classified as normal, first-level out-of-tolerance, and second-level out-of-tolerance.

8. The method for transferring industrial operation skills to embodied robots according to claim 1, characterized in that, S5 includes: Based on the displacement difference or end micro-displacement calculation results between the joint motor side and the output side, the transmission dead zone is identified by combining the pre-calibrated static friction identification threshold, joint drive current, command direction, equivalent angular displacement difference, output side angular velocity and duration. After the determination is established, an alternating compensation signal is superimposed on the corresponding joint. The frequency and amplitude of the alternating compensation signal are determined based on the structural modal test results and static friction identification results. The determination of dead zone release, compensation cancellation control, or compensation failure is made based on the difference in equivalent angular displacement after compensation, the angular velocity on the output side, the continuity of micro-displacement at the end, the change in contact force, and the change in current within the observation window.

9. A method for transferring industrial operation skills to embodied robots according to claim 1, characterized in that, S6 includes: When the contact force exceeds the process safety boundary corresponding to the current task, or the dead zone compensation fails and the contact force continues to increase, or the risk of continuous chatter is not eliminated after local parameter adjustment, stop advancing to the current position and retreat in the opposite direction of the current contact force vector; After the retreat is completed, clear the contact data fragments before the limit is exceeded, reconstruct the spatial relationship between the workpiece and the end, trigger the target area to be rescanned and reset the control parameters; When the target area offset is within the allowable range and the number of automatic retries has not reached the upper limit, the conservative retry parameter group is invoked. When the target area offset exceeds the allowable range or the number of automatic retries reaches the limit, the system enters the manual intervention state and outputs a process exception code.

10. A method for transferring industrial operation skills to embodied robots according to claim 9, characterized in that, After the retreat is completed, clear the contact data fragments before the limit was exceeded, reconstruct the spatial relationship between the workpiece and the end effector, trigger a rescan of the target area and reset the control parameters, including: After the retreat ends, the state is reset, the out-of-limit contact data fragments in the historical state buffer are cleared, the spatial relationship between the current workpiece position and the end position is re-established, the target area scan is triggered, and the contact control parameters, fusion weights and compensation signal states are restored to the safe initial settings. The synchronization state before the limit was exceeded, the contact difference evaluation results, the process of fusion weight change, the joint dead zone identification results, and the contact force direction at the moment of retreat triggering were traced back, and abnormal samples were recorded.

Images