Mechanical arm visual deep learning adaptive control system for complex working conditions

Abstract

The application relates to the technical field of mechanical arm vision control, and discloses a mechanical arm vision deep learning adaptive control system for complex working conditions, which comprises an asynchronous data alignment module, a coupled feature extraction module, an adaptive filtering module, a saccade active detection module and an impedance safety defense module. The absolute time stamp of the camera exposure midpoint is recorded to backtrack the actual linear speed of the mechanical arm, the alignment residual linear speed vector is calculated, and the observation noise covariance matrix of the filter is dynamically adjusted in combination with the high-frequency Shannon entropy of the image; microscopic saccade periodic oscillation and instantaneous zero-speed exposure are implemented in the visual blind area to explore the escape path; when the residual error is seriously out of limit, the stiffness and damping matrix are remodeled based on the error direction, and the system is switched to the admittance mode to execute passive safety slow stop. The application eliminates the calculation misplacement between high-frequency bottom-layer control and low-frequency vision reasoning, and endows the mechanical arm with active vision recovery capability and asymmetric physical compliance protection under extreme working conditions.

Description

Technical Field

[0001] This invention relates to the field of robotic arm vision control technology, specifically to a robotic arm vision deep learning adaptive control system for complex working conditions. Background Technology

[0002] Currently, industrial robotic arms are widely used in complex manufacturing and unstructured environments. Visual sensors combined with deep learning algorithms have become the core means of acquiring information about the external environment. Deep learning models can process massive amounts of image features, providing robotic arms with target recognition and pose prediction data. This vision control system enables robotic arms to possess a certain degree of autonomous operation capability. In welding, assembly, and painting processes, vision-guided dynamic pose adjustment is gradually replacing traditional fixed-path programming.

[0003] For the aforementioned vision control scenarios, conventional systems typically employ a cascaded control architecture. The bottom-level servo driver continuously reads joint states and executes position closure at a fixed high-frequency cycle. The upper-level computing unit processes image frames captured by the camera, outputs predicted coordinates using a network model, and then sends commands down to the bottom layer. Data transmission and fusion largely rely on conventional Kalman filtering algorithms to smooth the visual coordinates. When the robotic arm needs to cope with external contact forces, the system generally pre-sets globally fixed stiffness and damping parameters. The control module uses admittance control algorithms to ensure the robotic arm exhibits a consistent, compliant state in three-dimensional space.

[0004] However, actual operation under complex conditions has significant limitations. The large latency of visual inference and the high-frequency sampling of the underlying controller have inherent periodic differences. Directly comparing these two types of asynchronous data inevitably leads to calculation misalignment in the physical time dimension, resulting in severe distortion of motion deviation assessment. In addition, existing filtering mechanisms are mostly in a passive observation state. Once dust obstruction or strong light interference occurs on site, the quality of visual data continues to deteriorate, and the system lacks active means to explore the external environment to escape difficulties, which easily leads to the continuous accumulation of erroneous position commands in the visual blind zone. On the other hand, the underlying safety protection strategy is relatively simple when encountering extreme visual sudden errors. The globally symmetrical impedance parameters prevent the robotic arm from applying high stiffness resistance to specific directions where errors diverge sharply, making it difficult to simultaneously take into account directional interception of overstepping boundaries and non-directional absorption of collision energy, ultimately affecting the reliability of passive safety braking.

[0005] Therefore, this invention provides a vision-deep learning adaptive control system for robotic arms in complex working conditions to address the shortcomings of existing technologies. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a deep learning-based adaptive control system for robotic arms under complex working conditions. This system solves the problems of existing robotic arm vision control systems, such as deviation and distortion caused by spatiotemporal misalignment of high and low frequency data under complex interference conditions, lack of active recovery path in continuous visual blind spots, and inability of underlying physical protection to simultaneously address directional interception and flexible adaptation in extreme visual failures.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A vision-based deep learning adaptive control system for robotic arms in complex working conditions includes:

[0009] The asynchronous data alignment module records the absolute timestamp of the camera exposure midpoint and outputs visual prediction pose data, pushing the current actual linear velocity vector of the robotic arm into the first-in-first-out circular buffer.

[0010] The coupled feature extraction module backtracks the historical actual linear velocity vector based on the absolute timestamp of the camera exposure midpoint, calculates the aligned residual linear velocity vector, and outputs the high-frequency Shannon entropy of the image.

[0011] The adaptive filtering module calculates the dynamic observation noise covariance matrix based on the residual linear velocity amplitude generated by the high-frequency Shannon entropy of the image and the aligned residual linear velocity vector, and outputs the corrected pose control command to perform adaptive control on the robotic arm.

[0012] The scanning active detection module outputs a micro-scanning displacement vector trajectory command and a visual escape command to drive the robotic arm to adjust its motion trajectory when the trace of the dynamic observation noise covariance matrix exceeds a preset trace limit and the absolute value of the discrete-time first derivative of the high-frequency Shannon entropy of the image is less than a preset derivative threshold.

[0013] The impedance safety defense module, when the residual linear velocity amplitude exceeds the preset safety constraint threshold, performs passive safety slow-stop on the robotic arm based on the aligned residual linear velocity vector combined with the admittance control mode.

[0014] By adopting the above technical solutions, the physical time dimension calculation misalignment error caused by the difference in the sampling period of the visual sensor and the underlying kinematics of the robotic arm is eliminated by using an asynchronous data alignment module to record the equivalent exposure time of the camera and establish an absolute time axis for data addressing; a multi-dimensional system observation quality evaluation system is constructed by using a coupled feature extraction module to simultaneously calculate the high-frequency Shannon entropy of the image representing the optical degradation state of the external environment and the alignment residual linear velocity vector representing the degree of deviation between the visual deduced motion state and the actual motion state of the body; an adaptive filtering module is used to correlate the above features with the covariance inflation factor to achieve nonlinear smooth decay of the visual observation weight under complex working conditions; a scanning active detection module applies spatial periodic oscillation motion when it detects that the covariance matrix is ​​diverging and the image quality is no longer fluctuating, enabling the control system to have the active escape capability to explore the visual recovery path in occluded or strongly interfered areas; and an impedance safety defense module uses the directional gain matrix to reshape the asymmetric impedance parameters when the residual is severely exceeded, thereby achieving the technical effect of maintaining the stability of the robotic arm's motion trajectory and providing physical torque compliance protection under extreme visual abrupt changes.

[0015] Preferably, the asynchronous data alignment module executes the recording of the absolute timestamp of the camera exposure midpoint and outputs visual prediction pose data, and pushes the current actual linear velocity vector of the robotic arm into the first-in-first-out circular buffer, specifically including:

[0016] Establish an absolute time axis and continuously read the state data of the underlying robotic arm's motion joints according to a fixed high-frequency control cycle;

[0017] Calculate the current actual linear velocity vector of the robotic arm's end effector based on the stated state data and the forward kinematics model;

[0018] The current actual linear velocity vector is combined with the corresponding high-frequency timestamp data to form a motion state frame, and the motion state frame is pushed into the first-in-first-out circular buffer.

[0019] The visual prediction pose data is output by processing asynchronously input image frames from an industrial camera using a deep learning network model, and the equivalent center time of the image frame capturing external optical information is recorded as the absolute timestamp of the camera exposure midpoint.

[0020] By adopting the above technical solution, a spatiotemporal synchronization benchmark accurate to the midpoint of exposure is constructed, which solves the lag error caused by the inference time of deep learning models and the delay in image transmission.

[0021] Preferably, the coupled feature extraction module performs the calculation of the aligned residual linear velocity vector by backtracking the historical actual linear velocity vector based on the absolute timestamp of the camera exposure midpoint. This specifically includes:

[0022] Using the absolute timestamp of the camera exposure midpoint as the retrieval condition, the system addresses the first-in-first-out circular buffer and backtracks to read the historical actual linear velocity vector that strictly corresponds to the absolute timestamp of the camera exposure midpoint.

[0023] A time-different operation is performed on multiple continuously received visual prediction pose data to derive the visual derivation line velocity vector corresponding to the absolute timestamp of the camera exposure midpoint.

[0024] Calculate the vector difference between the visual derivation line velocity vector and the historical actual line velocity vector to generate the alignment residual line velocity vector that eliminates the calculation misalignment error.

[0025] By adopting the above technical solution, the visually inferred motion state and the actual physical motion state of the body are compared on a strictly aligned time axis to extract pure motion deviation features.

[0026] Preferably, the coupled feature extraction module performs the high-frequency Shannon entropy of the output image, specifically including:

[0027] Acquire a single frame image captured in parallel, and perform a two-dimensional fast Fourier transform on the spatial domain two-dimensional pixel matrix of the single frame image to transform the single frame image into a two-dimensional frequency domain space.

[0028] Extract high-frequency mask region data above the set cutoff frequency in the two-dimensional frequency domain space to construct the high-frequency power spectrum of the current image frame;

[0029] Normalization processing is performed on the high-frequency power spectral density data to generate a normalized high-frequency power spectral density;

[0030] The normalized high-frequency power spectral density is used in combination with a logarithmic function and a summation operation to calculate and output the high-frequency Shannon entropy of the image that quantifies the external optical degradation state.

[0031] By adopting the above technical solution, the uncertainty of image information is quantified by utilizing the principles of frequency domain transformation and information entropy, thereby achieving objective numerical calibration of the loss of high-frequency details in images caused by dust occlusion or strong light interference.

[0032] Preferably, the adaptive filtering module performs the calculation of the dynamic observation noise covariance matrix based on the residual linear velocity amplitude generated by the high-frequency Shannon entropy of the image and the aligned residual linear velocity vector, specifically including:

[0033] Extract the three-dimensional coordinate components of the aligned residual linear velocity vector and calculate the square root of the sum of squares of each coordinate component to convert it into a scalar value, thereby generating the residual linear velocity amplitude;

[0034] A modulation function containing a bidirectional exponential decay term is executed, the first term of which is used to calculate the covariance inflation factor caused by environmental optical degradation based on the high-frequency Shannon entropy of the image;

[0035] The second term of the modulation function is based on the extracted residual linear velocity amplitude, and the covariance inflation factor caused by the deviation between the visually derived state and the actual motion state of the body is calculated using the inverse exponential decay function.

[0036] The penalty gain of the first and second terms is superimposed to generate the dynamic observation noise covariance matrix based on the baseline observation noise covariance matrix.

[0037] By adopting the above technical solution, based on the dual modulation function of external optical characteristics and internal motion deviation characteristics, the observation noise matrix is ​​dynamically adjusted by superimposing penalty gain, thereby realizing the nonlinear adaptive shrinkage of Kalman observation weights.

[0038] Preferably, the adaptive filtering module executes the corrected pose control command to perform adaptive control on the robotic arm, specifically including:

[0039] The generated dynamic observation noise covariance matrix is ​​input into the internal extended Kalman filter as the observation noise parameter;

[0040] The current Kalman gain matrix is ​​calculated using the Kalman gain equation, which automatically reduces the weight of the dependence on the current visual prediction pose data.

[0041] By combining the reduced-weight visual observation data with the state prediction data based on the kinematic model, the pose state of the robotic arm end effector is smoothly estimated and the state recursion is calculated to generate the corrected pose control command.

[0042] The corrected pose control command is sent to the underlying servo driver to control the movement trajectory of the robotic arm.

[0043] By adopting the above technical solution, the weight of the internal positive kinematics model is automatically increased when visual confidence decreases, ensuring the continuity and smoothness of the control system command issuance.

[0044] Preferably, the active scanning detection module performs the following actions when the trace of the dynamic observation noise covariance matrix exceeds a preset trace limit and the absolute value of the discrete-time first derivative of the high-frequency Shannon entropy of the image is less than a preset derivative threshold:

[0045] The trace of the dynamic observation noise covariance matrix is ​​obtained by reading the dynamic observation noise covariance matrix in real time and calculating the sum of the main diagonal elements of the dynamic observation noise covariance matrix.

[0046] The high-frequency Shannon entropy sequence of the image for multiple consecutive control cycles is obtained, and the discrete-time first derivative is calculated using a numerical differentiation algorithm.

[0047] When it is determined that the trace of the dynamic observation noise covariance matrix exceeds the preset trace limit and the absolute value of the discrete-time first derivative of the high-frequency Shannon entropy of the image is less than the preset derivative threshold within multiple consecutive control cycles, a control cutoff command is issued to suspend the current position servo control.

[0048] By adopting the above technical solution, the trace of the covariance matrix is ​​used as a numerical evaluation index of the system state divergence. Combined with the differential change rate of the Shannon entropy of the image, the lingering state of environmental interference is judged, and the active detection mechanism is accurately triggered.

[0049] Preferably, after pausing the current position servo control, the scanning active detection module is further used for:

[0050] The microscopic scanning displacement vector trajectory command containing orthogonal periodic signals with different frequencies and amplitudes is applied to the underlying controller to drive the end effector of the robotic arm to perform spatial periodic oscillation motion with a preset amplitude.

[0051] During the oscillating motion, the theoretical instantaneous velocity vector calculated is continuously monitored. When the magnitude of the three-dimensional coordinate component drops to zero or below the preset velocity threshold, it is determined that the instantaneous zero-crossing point in three-dimensional space has been reached, and a hard trigger command is issued through the hardware pin to force the industrial camera to conduct a short exposure to acquire the image.

[0052] By comparing the high-frequency Shannon entropy of the image acquired through continuous discrete zero-speed point exposures, the visual escape command is output according to the negative gradient direction in the discrete space to reconstruct the subsequent control trajectory.

[0053] By adopting the above technical solution, hard-triggered exposure is implemented at the instantaneous zero speed point of periodic oscillation to avoid motion blur, and the robotic arm is guided to leave the visual blind zone according to the decreasing direction of the image information entropy gradient.

[0054] Preferably, the impedance safety defense module performs passive safety easing of the robotic arm based on the aligned residual linear velocity vector combined with the admittance control mode when the residual linear velocity amplitude exceeds the preset safety constraint threshold, specifically including:

[0055] The residual linear velocity amplitude is monitored in real time. When it is determined that the residual linear velocity amplitude exceeds the preset safety constraint threshold, a global safety back-off response mechanism is triggered and the coordinate data output of the adaptive filtering module is stopped.

[0056] Obtain the alignment residual linear velocity vector and perform a normalization operation on the alignment residual linear velocity vector to generate an error direction unit vector characterizing the error divergence direction;

[0057] The outer product of the error direction unit vector and the transpose of the error direction unit vector is calculated to construct the direction gain matrix. The direction gain matrix is ​​multiplied by the set stiffness penalty scalar and superimposed on the reference stiffness matrix to generate the plastic stiffness matrix.

[0058] By performing the same outer product recombination operation on the basic damping parameters using the error direction unit vector, a plastic damping matrix that strictly corresponds to the plastic stiffness matrix is ​​generated.

[0059] By adopting the above technical solution, the directional gain matrix is ​​generated based on the outer product operation of the error vector and the transpose matrix, thereby reshaping the reference stiffness and basic damping parameters and realizing the asymmetric allocation of control parameters in multidimensional space.

[0060] Preferably, after generating the plastic stiffness matrix and the plastic damping matrix, the impedance safety defense module is further used for:

[0061] The generation mode switching command is sent to the underlying controller to switch the operating mode from the position control mode to the admittance control mode;

[0062] In the admittance control mode, the actual sensing data of the underlying torque sensors installed on each motion joint or end effector of the robotic arm are read in real time.

[0063] The plastic stiffness matrix and plastic damping matrix are injected as new control parameters into the admittance control algorithm;

[0064] Under the combined effect of torque feedback data and reshaped impedance parameters, the robotic arm exhibits flexible compliance when facing the external environment, while displaying high stiffness and high damping in the specific direction of residual divergence, thereby achieving passive safe slow stop.

[0065] By adopting the above technical solution, when a serious visual inference error occurs, the robotic arm can exhibit a high impedance state in a specific direction where the error diverges sharply to resist overshooting, while maintaining flexible compliance in other directions to absorb collision energy.

[0066] This invention provides a vision-based deep learning adaptive control system for robotic arms operating in complex conditions. It offers the following advantages:

[0067] 1. This invention records the absolute timestamp of the camera exposure midpoint using an asynchronous data alignment module and addresses it in a first-in-first-out circular buffer, eliminating computational misalignment errors caused by period differences between the visual sensor inference delay and the high-frequency sampling of the robotic arm's underlying kinematics. Simultaneously, by combining the image's high-frequency Shannon entropy and the aligned residual linear velocity vector output by the coupled feature extraction module, the system can objectively evaluate the quality of visual data from two dimensions: the optical degradation state of the external environment and the degree of motion deviation derived internally. This multi-dimensional feature evaluation system based on precise time alignment avoids control misjudgments caused by single evaluation indicators under complex lighting and high-latency conditions, improving the system's accuracy in assessing the reliability of pose prediction.

[0068] 2. This invention combines passive weight adjustment with an active obstacle avoidance mechanism through an adaptive filtering module and a saccade active detection module. The system dynamically calculates the observation noise covariance matrix using a modulation function that includes gains with dual penalties for image features and motion deviations, smoothly reducing the visual observation weights as visual confidence decreases. When the trace of the covariance matrix exceeds the limit and the rate of change of the image Shannon entropy tends to stagnate, the system drives the robotic arm to perform microscopic saccade periodic oscillations and implements hard-triggered exposure at the instantaneous zero-velocity point. This mechanism enables the robotic arm to prevent the continuous accumulation of erroneous commands and to detect the surrounding environment without ambiguity when encountering continuous dust obstruction or strong interference blind zones, thereby reconstructing the trajectory along the negative gradient direction in space to restore normal visual servo control.

[0069] 3. This invention establishes a low-level physical protection mechanism for extreme visual abrupt changes through an impedance safety defense module. When the residual linear velocity amplitude exceeds the safety constraint threshold, the system reconstructs the directional gain matrix using the outer product operation of the error direction unit vector, thereby generating an asymmetric plastic stiffness matrix and plastic damping matrix, and switches the low-level drive to admittance control mode. This enables the robotic arm to exhibit high stiffness and high damping only in specific directions where the residual diverges sharply when a serious error occurs in visual inference, thus limiting out-of-bounds movement, while maintaining flexible compliance based on torque feedback in other spatial directions. This effectively absorbs collision energy when accidentally contacting environmental entities, achieving reliable passive safety slowing down. Attached Figure Description

[0070] Figure 1 This is a system framework diagram of the visual deep learning adaptive control system for robotic arms in complex working conditions according to the present invention.

[0071] Figure 2 This is a flowchart of the visual deep learning adaptive control method for robotic arms under complex working conditions according to the present invention.

[0072] Figure 3 This is a flowchart illustrating the asynchronous alignment of execution data according to the present invention;

[0073] Figure 4 This is a flowchart illustrating the extraction of coupling perturbation features according to the present invention;

[0074] Figure 5 This is a flowchart of the modulation adaptive covariance of the present invention;

[0075] Figure 6 This is a flowchart illustrating the active scanning detection process of the present invention;

[0076] Figure 7 The flowchart of the remodeling impedance security defense of the present invention is shown below;

[0077] Figure 8 This is a simulation diagram comparing the speed deviation of the adaptive control process of the present invention;

[0078] Figure 9 This is a comparison diagram of the contact forces of the impedance safety defense of the present invention. Detailed Implementation

[0079] The technical solutions in 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.

[0080] See attached document Figure 1 The present invention provides a vision deep learning adaptive control system for robotic arms for complex working conditions, which may include: an asynchronous data alignment module, a coupled feature extraction module, an adaptive filtering module, a scanning active detection module, and an impedance safety defense module.

[0081] The asynchronous data alignment module is configured at the front-end input layer of the control system to establish a global time synchronization reference and acquire underlying motion data. This module is equipped with a high-precision independent clock generator to establish a unified absolute time axis. Following a set fixed high-frequency control cycle, the asynchronous data alignment module continuously reads the status data of the underlying robotic arm's joints via the communication bus. This status data includes the current absolute rotation angle and angular velocity of each joint.

[0082] The asynchronous data alignment module integrates a forward kinematics calculation unit to calculate the actual linear velocity vector of the end effector based on the received state data. After calculation, the asynchronous data alignment module combines the actual linear velocity vector with the high-frequency timestamp data corresponding to the current moment and pushes it into a preset first-in-first-out circular buffer. The circular buffer ensures the temporal integrity of historical motion state data, facilitating backtracking and querying of historical data by subsequent modules.

[0083] The asynchronous data alignment module also establishes a data communication connection with an external industrial camera. When the industrial camera asynchronously transmits acquired image frames, the asynchronous data alignment module records the absolute timestamp of the camera exposure midpoint corresponding to that image frame. The image frame is then input into a deep learning network model for forward inference, outputting the visual prediction pose data of the robotic arm's end effector. By accurately recording the exposure timestamps of the image frames, this module achieves alignment and calibration of high-frequency robotic arm body states with low-frequency visual prediction poses in the time dimension.

[0084] The coupled feature extraction module is connected to the data output interface of the asynchronous data alignment module to extract various feature values ​​reflecting system operational deviations. The coupled feature extraction module obtains the absolute timestamp of the recorded camera exposure midpoint and uses this timestamp as a retrieval condition to backtrack and read the actual linear velocity vector strictly corresponding to that historical moment from the circular buffer. Simultaneously, this module performs temporal difference calculations using the received continuous visual prediction pose data to obtain the corresponding visually derived linear velocity vector.

[0085] The coupled feature extraction module calculates the difference between the visually derived linear velocity vector and the actual linear velocity vector obtained through backtracking, generating an aligned residual linear velocity vector. This aligned residual linear velocity vector objectively represents the spatial deviation between the deep learning network model output and the actual motion state of the robotic arm under a unified time reference, thereby eliminating calculation errors caused by asynchronous sampling of high- and low-frequency sensors.

[0086] The coupled feature extraction module also receives raw image frames acquired by an industrial camera in parallel and performs a two-dimensional fast Fourier transform on the image frame using the built-in signal processing unit. By extracting high-frequency mask region data above a set cutoff frequency in the two-dimensional frequency domain, a high-frequency power spectrum is constructed. After normalizing the high-frequency power spectral density data, the module calculates and outputs the high-frequency Shannon entropy of the image. The value of the high-frequency Shannon entropy directly quantifies the degree of disorder in the texture distribution within a specific high-frequency mask region of the current visual frame, serving as an objective indicator for evaluating ambient lighting and occlusion conditions.

[0087] The adaptive filtering module is connected to the coupled feature extraction module and is used to receive the high-frequency Shannon entropy and aligned residual linear velocity vector of the image and perform smooth state updates. During the observation update phase of the filter, the adaptive filtering module first calculates the L2 scalar of the aligned residual linear velocity vector and uses it as the residual linear velocity amplitude parameter.

[0088] The adaptive filtering module incorporates a dynamic calculation program for the observation noise covariance matrix. This program utilizes the high-frequency Shannon entropy of the input image and the residual linear velocity amplitude, combined with a predefined bidirectional exponential decay function, to generate the dynamic observation noise covariance matrix. The specific value of this matrix is ​​dynamically adjusted in real time based on the degree of degradation of the visual image and the degree of deviation in motion, used to quantify the unconfidence level of the current visual observation data.

[0089] The adaptive filtering module is internally configured with an extended Kalman filter. The dynamic observation noise covariance matrix calculated above is input into the extended Kalman filter as the observation noise parameter. The extended Kalman filter performs smooth estimation and recursive calculation on the pose state of the end effector based on the adjusted covariance matrix. This module outputs the corrected pose control command and sends it to the underlying servo driver to control the movement trajectory of the robotic arm.

[0090] The scanning active detection module connects to the adaptive filtering module and the industrial camera to guide the robotic arm in performing avoidance maneuvers when the optical environment deteriorates. The monitoring program within the scanning active detection module continuously reads the trace of the dynamic observation noise covariance matrix and the discrete-time derivative of the image's high-frequency Shannon entropy. When the trace of the dynamic observation noise covariance matrix exceeds a preset trace limit, and the absolute value of the first derivative of the image's high-frequency Shannon entropy is less than a preset derivative threshold, the module outputs a control cutoff command to pause the current position servo control.

[0091] The active scanning detection module then generates a microscopic scanning displacement vector trajectory command. This command contains orthogonal periodic signals with different frequencies and amplitudes. The active scanning detection module applies the microscopic scanning displacement vector trajectory as a local three-dimensional spatial position offset to the controller, driving the robotic arm's end effector to perform a spatial periodic oscillation motion of a preset amplitude at the current position.

[0092] During the execution of the microscopic scanned displacement vector trajectory, the computational unit of the scanned active detection module calculates the instantaneous zero-crossing point in three-dimensional space in real time. When the end effector reaches the instantaneous zero-crossing point, the module issues a hard trigger command through the controller hardware pin, forcing the industrial camera to perform a short exposure at that zero-velocity point and acquire an image. The scanned active detection module compares the high-frequency Shannon entropy of the images acquired through exposures at consecutive discrete zero-velocity points, calculates the bias according to the negative gradient direction in discrete space, and outputs a bias command to reconstruct the subsequent control trajectory.

[0093] The impedance safety defense module is connected to the coupling feature extraction module and the adaptive filtering module to perform passive takeover and resistance control when the system output is abnormal. The impedance safety defense module monitors the numerical change of the residual linear velocity amplitude in real time. When it is determined that the residual linear velocity amplitude exceeds the preset safety constraint threshold, a global safety back-off response mechanism is triggered, and a defense takeover signal is sent to the control system.

[0094] Upon receiving the defense takeover signal, the impedance safety defense module stops the coordinate data output of the adaptive filtering module. It then acquires the aligned residual linear velocity vector and performs a normalization operation on it to generate an error direction unit vector. Based on this error direction unit vector, the impedance safety defense module performs anisotropic reorganization on the robot arm's basic stiffness parameters in Cartesian space, calculating and generating the shaping stiffness matrix.

[0095] The impedance safety defense module uses a unit vector in the same error direction to perform the same recombination operation on the basic damping parameters, generating the corresponding plastic damping matrix. Simultaneously, it switches the control mode of the underlying controller from position mode to admittance control mode. In admittance control mode, the impedance safety defense module relies on the sensing data from the underlying torque sensor, combined with the plastic stiffness matrix and plastic damping matrix, to perform passive safety easing of the robotic arm.

[0096] See attached document Figure 2 The system workflow of the present invention is as follows, and may include:

[0097] Step S100: Perform asynchronous data alignment.

[0098] The asynchronous data alignment module establishes an absolute time axis based on an internal independent clock generator and continuously reads the status data of the underlying robotic arm's motion joints according to a fixed high-frequency control cycle. This status data includes the current absolute rotation angle and rotational angular velocity of each motion joint.

[0099] The asynchronous data alignment module uses the forward kinematics calculation unit to calculate the actual linear velocity vector of the robotic arm's end effector. After the calculation is completed, the asynchronous data alignment module combines the actual linear velocity vector with the current high-frequency timestamp data and pushes it into a first-in-first-out circular buffer for storage, thereby maintaining the temporal integrity of historical motion state data.

[0100] The asynchronous data alignment module receives image frames from an external industrial camera and simultaneously records the absolute timestamp of the camera's exposure midpoint corresponding to the image frame. Subsequently, the asynchronous data alignment module inputs the image frames into a deep learning network model for forward inference and outputs visually predicted pose data, thereby achieving high-precision alignment between the actual motion state of the subject and the visually predicted pose in the time dimension.

[0101] Step S200: Extract coupling perturbation features.

[0102] The coupled feature extraction module uses the previously recorded absolute timestamp of the camera exposure midpoint as the retrieval condition to backtrack and read the actual linear velocity vector that strictly corresponds to that historical moment from the circular buffer. Simultaneously, the coupled feature extraction module performs temporal difference operations using continuously received visually predicted pose data to generate a visually deduced linear velocity vector.

[0103] The coupled feature extraction module calculates the difference between the visually derived linear velocity vector and the actual linear velocity vector obtained through backtracking, generating an aligned residual linear velocity vector. This aligned residual linear velocity vector objectively represents the spatial deviation between the output of the deep learning network model and the actual motion state of the robotic arm under a unified time reference.

[0104] The coupled feature extraction module uses its internal signal processing unit to perform a two-dimensional fast Fourier transform on the synchronously received image frames, extracting high-frequency mask region data above a set cutoff frequency in the two-dimensional frequency domain to construct a high-frequency power spectrum. After normalizing the high-frequency power spectral density data, the coupled feature extraction module calculates and outputs the high-frequency Shannon entropy of the image.

[0105] Step S300: Modulate adaptive covariance.

[0106] The adaptive filtering module acquires the high-frequency Shannon entropy and aligned residual linear velocity vector of the image output by the coupled feature extraction module. During the observation update phase of the adaptive filtering module, it calculates the L2 scalar of the aligned residual linear velocity vector and determines the result of this scalar as the residual linear velocity amplitude parameter.

[0107] The adaptive filtering module calls its internally defined bidirectional exponential decay function, using the acquired high-frequency Shannon entropy of the image and the calculated residual linear velocity amplitude parameters to perform calculations and generate a dynamic observation noise covariance matrix. This dynamic observation noise covariance matrix directly quantifies the non-confidence level of the current visual observation data during system update.

[0108] The adaptive filtering module inputs the calculated dynamic observation noise covariance matrix into the internal extended Kalman filter as the observation noise parameter. Based on this dynamic parameter, the extended Kalman filter performs smooth estimation and recursive calculation of the pose state of the robotic arm's end effector, and finally outputs the corrected pose control command.

[0109] Step S400: Perform active scanning detection.

[0110] The scanning active detection module continuously reads the trace of the dynamic observation noise covariance matrix in the extended Kalman filter and the discrete-time derivative of the high-frequency Shannon entropy of the image output by the coupled feature extraction module. When the trace of the dynamic observation noise covariance matrix exceeds a preset trace limit and the absolute value of the first derivative of the high-frequency Shannon entropy of the image is less than a preset derivative threshold, the scanning active detection module issues a control cutoff command to pause the current position servo.

[0111] The active scanning detection module generates a microscopic scanning displacement vector trajectory command, which is applied to the controller as a local three-dimensional spatial position offset to drive the robotic arm's end effector to perform a spatial periodic oscillating motion of a preset amplitude. During the oscillating motion, the active scanning detection module calculates the instantaneous zero-crossing point in three-dimensional space in real time.

[0112] When the robotic arm's end effector detects an instantaneous zero-crossing, the scanning active detection module sends a hard trigger command via the controller's hardware pins, forcing the industrial camera to perform a short exposure at that zero-velocity point and acquire an image. The scanning active detection module compares the high-frequency Shannon entropy of the images acquired at different discrete zero-velocity points and reconstructs the control trajectory by outputting a bias command according to the negative gradient direction in the discrete space.

[0113] Step S500: Reshape impedance for safety protection.

[0114] The impedance safety defense module monitors the residual linear velocity amplitude calculated by the adaptive filtering module in real time. When the residual linear velocity amplitude is determined to exceed the preset safety constraint threshold, the impedance safety defense module triggers a global safety back-off response mechanism, sends a defense takeover signal to the control system, and stops the coordinate data output of the adaptive filtering module.

[0115] The impedance safety defense module acquires the aligned residual linear velocity vector and performs a normalization operation on it to generate an error direction unit vector. Using this error direction unit vector, the impedance safety defense module performs anisotropic reorganization on the robot arm's basic stiffness parameters in Cartesian space, calculating and generating the shaping stiffness matrix.

[0116] The impedance safety defense module uses the same error direction unit vector to perform an equivalent recombination operation on the basic damping parameters to generate a plastic damping matrix. Subsequently, the impedance safety defense module switches the underlying control mode to admittance control mode, and combines the sensing data from the underlying torque sensor, the plastic stiffness matrix, and the plastic damping matrix to perform passive safety easing of the robotic arm.

[0117] See attached document Figure 3 Sub-step S101: Establish a global high-precision system clock for the controller to continuously collect the state data of the underlying robotic arm joints with a fixed high-frequency control cycle.

[0118] The asynchronous data alignment module is configured in the front-end input layer of the control system. This module contains a high-precision independent clock generator to establish a global time synchronization reference and define a unified absolute time axis. The establishment of this absolute time axis provides a unified time measurement reference for data fusion from heterogeneous sensors throughout the control system, eliminating the cumulative drift error caused by independent clock sources from different sensors.

[0119] After the absolute timeline is established, the asynchronous data alignment module begins operation according to a set fixed high-frequency control cycle. The asynchronous data alignment module continuously reads the motion status information of the underlying robotic arm via the communication bus. The communication bus connects the asynchronous data alignment module to the robotic arm's underlying servo controller, ensuring low latency and high throughput data transmission.

[0120] The asynchronous data alignment module reads motion state information, specifically the state data of the underlying robotic arm's joints. This state data includes the current actual absolute rotation angle and actual angular velocity of each joint. This data directly reflects the robotic arm's true spatial configuration and motion trend within the current high-frequency control cycle, serving as the input for subsequent kinematic calculations.

[0121] Sub-step S102: Calculate the current actual linear velocity vector of the robotic arm end effector based on the forward kinematics model, and push the calculation result along with the current high-frequency timestamp into the circular buffer.

[0122] The asynchronous data alignment module integrates a forward kinematics calculation unit. After acquiring state data including the current actual absolute rotation angle and actual rotational angular velocity, the forward kinematics calculation unit establishes a forward kinematics model based on preset robotic arm link parameters and geometric configuration. The forward kinematics model describes the mapping relationship between joint space motion variables and robotic arm end effector motion variables in Cartesian space.

[0123] The forward kinematics calculation unit uses the forward kinematics model and the received state data to calculate the current actual linear velocity vector of the robotic arm's end effector. Specifically, the forward kinematics calculation unit obtains the current actual linear velocity vector through formula calculation.

[0124] The calculation formula is as follows: ;

[0125] in, This is the current actual linear velocity vector, which represents the actual moving speed of the robotic arm's end effector within the current high-frequency control cycle, in m / s. The basic translation Jacobian matrix represents the kinematic mapping relationship matrix from joint space velocity to Cartesian space linear velocity under the current robot arm configuration; This is the current joint angle vector, which represents the actual absolute rotation angle of each joint of the robotic arm at the current moment, in rad. This is the current joint angular velocity vector, which represents the actual rotational angular velocity of each joint of the robotic arm at the current moment, in rad / s.

[0126] After the current actual linear velocity vector is calculated, the asynchronous data alignment module obtains the current high-frequency timestamp data generated by the system that establishes the absolute time axis. The asynchronous data alignment module then strictly pairs and combines the calculated current actual linear velocity vector with the current high-frequency timestamp data to form a motion state frame with a time tag.

[0127] Subsequently, the asynchronous data alignment module pushes the motion state frame into a preset first-in-first-out circular buffer. The circular buffer uses a fixed-length queue structure to store historical motion state data within a set time window. This storage mechanism ensures the temporal integrity of the motion state data, enabling subsequent modules to accurately backtrack and query historical state data based on a specific absolute timestamp.

[0128] Sub-step S103: Receive image frames asynchronously transmitted from an industrial camera, record the absolute timestamp of the camera exposure midpoint corresponding to the image frame, and output visual prediction pose data through a deep learning network model.

[0129] In addition to connecting to the underlying servo controller, the asynchronous data alignment module also establishes a data communication connection with an external industrial camera. The industrial camera continuously acquires environmental images according to its own sensor frame rate and working cycle, and transmits the acquired image frames to the asynchronous data alignment module asynchronously. The acquisition frequency of the industrial camera is lower than the update frequency of the underlying state data, forming a high- and low-frequency dual-source data input in the system.

[0130] When an industrial camera asynchronously transmits a single image frame, the hardware-triggered recording circuit of the asynchronous data alignment module captures the data arrival event. Based on the established absolute timeline, the asynchronous data alignment module accurately records the absolute timestamp of the camera exposure midpoint corresponding to that image frame. This absolute timestamp indicates the equivalent center time at which the image frame captures external optical information, serving as the benchmark for resolving time misalignments between visually derived data and the body's motion state data.

[0131] After recording the absolute timestamp of the camera exposure midpoint, the asynchronous data alignment module inputs the received image frames into the internally loaded deep learning network model. The deep learning network model extracts visual features from the image frames based on pre-trained weights and performs forward inference operations. After calculation by the deep learning network model, it outputs the visual predicted pose data of the robotic arm's end effector. By accurately recording the absolute timestamp of the camera exposure midpoint and combining it with a first-in-first-out (FIFO) circular buffer storage mechanism, the asynchronous data alignment module achieves high-precision alignment and calibration of the frequently updated robotic arm body state data and the low-frequency input visual predicted pose data in the absolute time dimension.

[0132] See attached document Figure 4Sub-step S201: Based on the recorded absolute timestamp of the camera exposure midpoint, backtrack the corresponding historical actual linear velocity vector from the circular buffer and calculate the alignment residual linear velocity vector between it and the visually derived linear velocity vector.

[0133] The coupled feature extraction module is connected to the data output interface of the asynchronous data alignment module in the system input layer, and is used to extract various feature values ​​reflecting the operational deviations of the control system. The coupled feature extraction module first obtains the absolute timestamp of the camera exposure midpoint corresponding to the single-frame image recorded by the asynchronous data alignment module. Then, using this absolute timestamp as the search criterion, the coupled feature extraction module performs time matching and data addressing within a first-in-first-out circular buffer.

[0134] Through time addressing, the coupled feature extraction module backtracks from the circular buffer to read the historical actual linear velocity vectors that strictly correspond to the exposure moment. Simultaneously, the coupled feature extraction module receives continuous visual prediction pose data output by the deep learning network model. The built-in processing unit performs temporal difference operations on the multiple continuously received visual prediction pose data to derive the visually derived linear velocity vector corresponding to the exposure moment.

[0135] After obtaining the two velocity vectors aligned under a unified time reference, the coupled feature extraction module calculates the vector difference between the visually derived linear velocity vector and the historically obtained actual linear velocity vector, generating an aligned residual linear velocity vector. The process of calculating the aligned residual linear velocity vector eliminates the calculation misalignment error caused by the asynchronous sampling of high- and low-frequency sensors.

[0136] The calculation formula is as follows:

[0137] ;

[0138] in, To align the residual linear velocity vector, it represents the spatial deviation between the velocity derived by the visual algorithm and the actual velocity of the body under the exposure time reference, in m / s; is the visual derivation line velocity vector, which represents the movement speed obtained by time difference based on continuous visual prediction of pose, and the unit is m / s; This is the historical actual linear velocity vector, which represents the actual moving speed read back from the circular buffer and strictly corresponding to the camera exposure time, in m / s.

[0139] Sub-step S202: Parallel capture of image frames, perform two-dimensional fast Fourier transform to construct high-frequency power spectrum, and extract data from the corresponding mask area.

[0140] While performing velocity residual feature extraction, the coupled feature extraction module receives raw image frame data continuously acquired from the industrial camera in parallel. The coupled feature extraction module is equipped with a signal processing unit. Upon receiving the raw image frame, this unit does not directly extract features from the spatial domain pixels, but instead sends them to a frequency domain transformation channel. This processing method transforms spatial light and shadow and texture variations into frequency domain energy distribution features.

[0141] The signal processing unit performs a two-dimensional fast Fourier transform (FFT) operation on the spatial domain two-dimensional pixel matrix of the captured single-frame image. Through the expansion of orthogonal basis functions, the two-dimensional FFT transforms the image data from the spatial coordinate system to the two-dimensional frequency domain. In the transformed two-dimensional frequency domain, the low-frequency components of the image are concentrated in the central region of the frequency plane, while the high-frequency components reflecting edges and noise are distributed in regions far from the center.

[0142] To assess image degradation, the signal processing unit sets a cutoff frequency threshold in the two-dimensional frequency domain. The coupled feature extraction module constructs a high-frequency mask covering the high-frequency region based on this cutoff frequency. Using this high-frequency mask, the module extracts high-frequency mask region data above the cutoff frequency in the two-dimensional frequency domain and constructs the high-frequency power spectrum of the current image frame accordingly. The high-frequency power spectrum characterizes the energy intensity of high-frequency information in the image.

[0143] Sub-step S203: Based on the normalized high-frequency power spectral density, calculate the high-frequency Shannon entropy of the image to quantify the degree of visual degradation.

[0144] To meet the probabilistic computation requirements of information entropy, the coupled feature extraction module performs normalization processing on the constructed high-frequency power spectral density data. The built-in normalization program traverses the discrete frequency coordinates within the high-frequency mask region and calculates the sum of the power of all high-frequency components within the high-frequency mask region. Subsequently, the power density of each independent high-frequency component is divided by this sum to generate the corresponding normalized high-frequency power spectral density.

[0145] After obtaining the normalized high-frequency power spectral density of all high-frequency components, the coupled feature extraction module performs calculations based on information entropy theory. The computation unit uses the normalized values ​​as the probability distribution of high-frequency information occurrences, and combines logarithmic functions with cumulative summation operations to calculate the high-frequency Shannon entropy of the output image.

[0146] The calculation formula is as follows:

[0147] ;

[0148] in, is the high-frequency Shannon entropy of the image, which represents the degree of disorder in the texture distribution of a visual frame within a specific high-frequency mask region, and its value ranges from 0 to 10; It is a set of high-frequency mask indices, which represents the set of coordinates of frequency components above the cutoff frequency in the two-dimensional frequency domain space; To normalize the high-frequency power spectral density, which represents the first... The ratio of the power of each high-frequency component to the total power in the mask region, with a value greater than 0 and less than or equal to 1.

[0149] The output value of the high-frequency Shannon entropy of the image serves as an evaluation index for quantifying the external optical environment state. When the control system is under dust cover or illumination variation conditions, the high-frequency noise introduced by image acquisition causes the high-frequency energy distribution to become discrete and disordered, thereby leading to an increase in the high-frequency Shannon entropy value of the image. The calculation result of this index serves as the data basis for subsequent filter state updates and covariance matrix dynamic modulation.

[0150] See attached document Figure 5 Sub-step S301: In the filter update stage, obtain the high-frequency Shannon entropy of the image output in step S200 and the L2 norm scalar of the aligned residual linear velocity vector.

[0151] The adaptive filtering module establishes a data communication connection with the coupled feature extraction module at the front end of the control system. When the control system's operating cycle progresses to the filter's observation and update phase, the adaptive filtering module initiates its internal data acquisition program. This program continuously reads the various feature evaluation indicators generated by the coupled feature extraction module in the previous processing phase through the data transmission interface.

[0152] The data received by the adaptive filtering module specifically includes the image's high-frequency Shannon entropy and the alignment residual linear velocity vector. The image's high-frequency Shannon entropy, as an objective numerical indicator, directly reflects the high-frequency noise distribution of the current input visual image in the frequency domain. The alignment residual linear velocity vector characterizes the spatial deviation between the deep learning network model's output and the actual motion state of the robotic arm under a unified time reference.

[0153] After successfully acquiring the aforementioned feature data, the mathematical calculation unit within the adaptive filtering module extracts the three-dimensional coordinate components of the aligned residual linear velocity vector. The mathematical calculation unit performs a L2 operation on these three-dimensional coordinate components, converting the vector-form deviation data into a single scalar value by calculating the square root of the sum of the squares of each coordinate component. This scalar value is defined as the residual linear velocity amplitude, used to accurately quantify the absolute spatial difference between the visually derived velocity and the actual velocity of the entity.

[0154] Sub-step S302: Based on the degree of deviation between the actual motion and the optics, calculate the dynamic observation noise covariance matrix using a modulation function that includes a bidirectional exponential attenuation term.

[0155] The adaptive filtering module is pre-configured with a dynamic calculation program for the observation noise covariance matrix and a benchmark observation noise covariance matrix. The benchmark observation noise covariance matrix records the inherent noise level of the sensor under ideal and interference-free conditions. Based on the aforementioned benchmark observation noise covariance matrix, the dynamic calculation program uses real-time acquired environmental and state characteristic data to dynamically quantify and adjust the non-confidence of the visual observation data currently being processed by the filter.

[0156] The dynamic calculation program executes a modulation function containing a bidirectional exponential decay term. The first term of this modulation function calculates the difference between the high-frequency Shannon entropy of the input image and a set baseline entropy threshold, and, combined with an exponential nonlinear amplification mechanism, calculates the covariance inflation factor caused by environmental optical degradation. The second term of this modulation function, based on the extracted residual linear velocity amplitude, uses an inverse exponential decay function to calculate the covariance inflation factor caused by the divergence between the visually derived state and the actual motion state of the body.

[0157] By superimposing the penalty gain from image frequency domain degradation and the penalty gain from actual motion deviation, the dynamic calculation program outputs the final covariance adjustment result. The calculation formula is as follows:

[0158] ;

[0159] in, The dynamic observation noise covariance matrix represents the non-confidence evaluation matrix of the filter on the visual observation results in the current update phase. The baseline observation noise covariance matrix represents the fixed sensor noise covariance configuration under ideal, interference-free conditions. For entropy feedback weight coefficient, it represents the penalty gain parameter for the increase of covariance matrix due to the phenomenon of image frequency domain degradation, and the value ranges from 1 to 100; is the entropy exponential decay constant, which represents the nonlinear rate parameter of covariance expansion after high-frequency entropy exceeds the baseline; is the high-frequency Shannon entropy of the image, which represents the degree of disorder in the texture distribution of a visual frame within a specific high-frequency mask region, and its value ranges from 0 to 10; The baseline entropy threshold represents the statistically average high-frequency Shannon entropy scalar calculated under normal and stable operating conditions, with a value ranging from 0 to 10. For residual feedback weighting coefficients, which represent the penalty gain parameter for the expansion of the covariance matrix due to the actual deviation of motion, the value ranges from 1 to 100; is the residual exponential decay constant, which represents the nonlinear rate parameter of the residual magnitude response; The residual linear velocity amplitude represents the scalar length of the aligned residual linear velocity vector in three-dimensional space, with units of m / s.

[0160] Sub-step S303: Substitute the dynamic observation noise covariance matrix into the adaptive extended Kalman filter to perform smooth estimation and state recursion for the pose vector.

[0161] The adaptive filtering module contains an adaptive extended Kalman filter algorithm entity. The control logic of the adaptive filtering module uses the dynamic observation noise covariance matrix calculated in sub-step S302 as the current observation noise parameter and directly inputs it into the adaptive extended Kalman filter. This data injection operation replaces the filter's default static noise covariance parameter, enabling the filter algorithm's parameter configuration to map in real time the degree of optical degradation in the external environment and the actual operational deviation of the control system.

[0162] Upon receiving the updated dynamic observation noise covariance matrix, the adaptive extended Kalman filter calculates the current Kalman gain matrix according to the Kalman gain equation. Since the value of the dynamic observation noise covariance matrix expands with the increase of the image's high-frequency Shannon entropy and residual linear velocity amplitude, the calculated Kalman gain will decrease accordingly. This mathematical relationship allows the filter to automatically reduce its reliance on poor-quality visual prediction pose data during state updates.

[0163] By combining reduced-weight visual observation data with kinematic model-based state prediction data, an adaptive extended Kalman filter performs smooth estimation and state recursion calculations on the pose state of the robotic arm's end effector. After the recursion calculation is completed, the adaptive filtering module outputs a corrected pose control command. This corrected pose control command is then continuously sent to the underlying servo driver via the control system's communication bus to precisely control the robotic arm's actual motion trajectory under complex working conditions.

[0164] See attached document Figure 6 Sub-step S401: Continuously monitor the system status. When the trace of the dynamic observation noise covariance matrix exceeds the preset trace limit and the first derivative of the high-frequency entropy approaches zero, disconnect the position servo and inject the microscopic scanning displacement vector trajectory into the end of the robotic arm.

[0165] The active scanning detection module establishes a communication connection with the adaptive filtering module to continuously read the state data during the filtering process. Specifically, the monitoring program inside the active scanning detection module acquires the dynamic observation noise covariance matrix in real time and calculates the sum of the main diagonal elements of this matrix, i.e., the trace of the dynamic observation noise covariance matrix. Simultaneously, this module acquires the high-frequency Shannon entropy sequence of the image output by the coupled feature extraction module and uses a numerical differentiation algorithm to calculate its discrete-time first derivative, thereby objectively quantifying the changing trend of the environmental optical state.

[0166] The scanning active detection module has preset limits for determining the system state. When the monitoring program determines that the trace of the currently calculated dynamic observation noise covariance matrix exceeds the preset limit, and the absolute value of the first derivative of the high-frequency Shannon entropy of the image is less than a preset derivative threshold for multiple consecutive control cycles, the system is determined to be in a severe and stable optical blind zone or occlusion state. Under this condition, the scanning active detection module sends a control cut-off command to the lower level of the control system, suspending the current position servo control of the robotic arm and preventing the accumulation of errors based on low-confidence visual data.

[0167] After the position servo control is disconnected, the internal trajectory generator of the scanning active detection module generates a micro-scanning displacement vector trajectory command. This command is composed of the superposition of orthogonal periodic signals with different frequencies and amplitudes. The scanning active detection module applies the micro-scanning displacement vector trajectory as a local three-dimensional spatial position offset to the underlying controller, driving the robotic arm's end effector to perform a spatial periodic oscillation motion of a preset amplitude at the current dwell position, thereby actively changing the camera's optical observation angle. The calculation formula is as follows:

[0168] ;

[0169] in, The microscopic scanning displacement vector represents the local three-dimensional spatial position offset actively injected by the underlying controller to the end effector of the robotic arm, with units of meters (m). The X-axis disturbance amplitude represents the displacement limit of the micro-motion scanning trajectory in the X-axis direction, with a value ranging from 0.001 to 0.005, and the unit is m; The value is the Y-axis disturbance amplitude, which represents the displacement limit of the micro-motion scanning trajectory in the Y-axis direction. The value ranges from 0.001 to 0.005, and the unit is m. The X-axis micro-motion angular frequency represents the periodic oscillation rate of the micro-motion scanning command in the X-axis direction, and its unit is rad / s; The micro-motion angular frequency along the Y-axis represents the periodic oscillation rate of the micro-motion scanning command in the Y-axis direction, and its unit is rad / s; This is the current absolute timestamp, representing the absolute time of the high-precision hardware clock of the system upon which the micro-motion trajectory function evaluation depends, in seconds; The orthogonal phase difference represents the initial time phase offset between the periodic signal functions of the X-axis and Y-axis, and its unit is rad. This refers to the matrix transpose or vector transpose in mathematical operations.

[0170] Sub-step S402: Calculate the instantaneous zero-crossing point of the three-dimensional space of the microscopic scanning displacement vector trajectory in real time, and issue a hard trigger command through the controller hardware pin to force the industrial camera to perform a short exposure at the instantaneous zero speed point.

[0171] During the spatial oscillation motion of the robotic arm's end effector based on the microscopic scan displacement vector trajectory, the mathematical operation unit of the scan active detection module performs analytical calculations on the kinematic derivative of the position bias function. This mathematical operation unit obtains the theoretical instantaneous velocity vector of the robotic arm's end effector on the oscillation trajectory by calculating the first-order time derivative of the microscopic scan displacement vector with respect to the current absolute timestamp.

[0172] The mathematical operation unit continuously monitors the magnitude of the calculated theoretical instantaneous velocity vector. When the magnitude of the three-dimensional coordinate component of this instantaneous velocity vector drops to zero or below a preset velocity threshold, the scanning active detection module determines that the end effector of the robotic arm has reached the instantaneous zero-crossing point of the oscillation trajectory in three-dimensional space. At these instantaneous zero-crossing points, the robotic arm is in a very brief state of relative stillness, and the structural vibration disturbance of the robotic arm body reaches a minimum.

[0173] Once the instantaneous zero-crossing point is detected, the scanning active detection module immediately sends a high-level hard trigger command through a dedicated hardware pin on the underlying controller. This hard trigger command is transmitted directly to the external trigger interface of the industrial camera via an electrical connection cable, forcing the industrial camera to perform a short exposure at the zero-speed point on the order of microseconds. Through this hardware interrupt-based forced exposure mechanism, the system can acquire a clear, original image frame without motion blur at the moment when the robotic arm's movement is at its most stable.

[0174] Sub-step S403: Compare the high-frequency Shannon entropy of the image acquired by discrete zero-speed point exposure, and reconstruct the control trajectory by outputting visual escape instructions according to the negative gradient direction of discrete space.

[0175] After the industrial camera completes a short exposure and acquires images, it asynchronously transmits the acquired image frames to the coupled feature extraction module of the control system. The coupled feature extraction module performs Fast Fourier Transform and frequency domain analysis on these image frames acquired at discrete instantaneous zero-crossing points according to a preset signal processing program, calculating and outputting the corresponding high-frequency Shannon entropy of each image. The scanning active detection module receives and stores this series of high-frequency Shannon entropy data calculated at discrete zero-crossing points, along with the corresponding three-dimensional spatial coordinates at the time of acquisition.

[0176] The scanning active detection module constructs a local spatial optical environment quality gradient field using stored discrete spatial coordinates and corresponding high-frequency Shannon entropy values ​​of the image. The gradient calculation program of this module performs spatial difference operations on the high-frequency Shannon entropy of the image at adjacent zero-crossing points. By comparing the magnitudes of the high-frequency entropy at different spatial locations, the rate of change of values ​​in various spatial directions is calculated, thereby determining the spatial direction that causes the fastest decrease in the image's high-frequency Shannon entropy; this direction is defined as the discrete spatial negative gradient direction.

[0177] After determining the negative gradient direction in the discrete space, the scanning active detection module calculates the escape bias along this direction and outputs a visual escape command containing this bias. The visual escape command is sent to the lower-level controller, guiding the robotic arm's end effector to gradually escape the current high-entropy occlusion region. Based on this, the control system's trajectory planning unit reconstructs the subsequent control trajectory using the new coordinate pose, restoring the normal position servo mode, allowing the robotic arm to continue performing its predetermined task under improved optical vision.

[0178] See attached document Figure 7 Sub-step S501: During system operation, when the residual linear velocity amplitude is detected to exceed the preset constraint threshold, a global safety rollback response is triggered.

[0179] The impedance safety defense module establishes and maintains data communication connections with the coupling feature extraction module and adaptive filtering module within the control system. During the continuous operation cycle of the control system, the monitoring program configured within the impedance safety defense module is activated and executed periodically. This monitoring program reads the residual linear velocity amplitude calculated and output by the adaptive filtering module during the pre-processing stage in real time through the communication interface.

[0180] The acquired residual linear velocity amplitude, as an objective scalar data point, accurately reflects the absolute spatial deviation between the current prediction result of the visual deep learning network and the actual motion state of the robotic arm. The impedance safety defense module has pre-allocated storage space to save the set safety constraint thresholds. These safety constraint thresholds are pre-calculated and fixed numerical standards based on the kinematic limits of the robotic arm and the safety tolerance boundaries of the working environment.

[0181] In each control cycle, the monitoring program rigorously compares the real-time residual linear velocity amplitude with the aforementioned stored safety constraint threshold. When the monitoring program determines that the real-time residual linear velocity amplitude exceeds the preset safety constraint threshold, the system state is deemed abnormal, indicating that the visually predicted trajectory has deviated from the actual safe operating range. Based on this determination, the impedance safety defense module immediately triggers the global safety backoff response mechanism, generating and sending a defense takeover signal to the main control logic unit of the control system, taking over the transmission channel of the underlying control commands.

[0182] Sub-step S502: Disable filter coordinate output, normalize the aligned residual linear velocity vector, and based on the calculated error direction unit vector, perform anisotropic reorganization of Cartesian space stiffness on the controller to generate a plastic stiffness matrix.

[0183] After successfully sending the defensive takeover signal, the impedance safety defense module cuts off the error accumulation path. The impedance safety defense module sends a stop command to the adaptive filtering module, blocking the adaptive filtering module from outputting smoothed, estimated pose coordinate data to the control system backend. This operation forces the system to discard the current low-confidence visual guidance trajectory, preventing the underlying servo drivers from receiving and executing erroneous motion commands.

[0184] After the coordinate output cutoff operation is completed, the impedance security defense module retrieves the alignment residual linear velocity vector corresponding to the current control cycle from the coupling feature extraction module via the communication interface. The vector operation unit inside the impedance security defense module receives this alignment residual linear velocity vector and performs a mathematical normalization operation on it. By dividing the three-dimensional components of the vector by the norm 2 of the vector, the operation unit eliminates the original velocity magnitude attribute of the vector and outputs a unit vector of error direction that simply represents the direction of error divergence.

[0185] The impedance safety defense module utilizes the generated error direction unit vector to perform anisotropic reconstruction calculations on the reference stiffness matrix preset within the underlying controller. This module constructs a directional gain matrix by calculating the outer product of the error direction unit vector and its transpose. This directional gain matrix is ​​then multiplied by a set stiffness penalty scalar and superimposed onto the basic three-dimensional symmetric stiffness matrix. In this way, the system can unidirectionally increase the resistance coefficient in a specific error divergence direction, generating a plastic stiffness matrix. The calculation formula is as follows:

[0186] ;

[0187] in, The shaping stiffness matrix represents the dynamic stiffness parameter matrix of the robotic arm in three-dimensional Cartesian space under admittance control mode, which has anisotropic resistance characteristics. The unit is N / m. The reference stiffness matrix represents the three-dimensional basic symmetric stiffness parameter matrix of the robotic arm under normal compliant mode, with units of N / m. For stiffness penalty scalar, it represents the resistance strength gain parameter added in the direction of visual error guidance, with a value range of 1000 to 10000, and the unit is N / m; Let be the unit vector of the error direction, which represents the normalized projection direction vector of the alignment residual linear velocity vector in three-dimensional Cartesian space; This refers to the matrix transpose or vector transpose in mathematical operations.

[0188] Sub-step S503: Synchronously generate the shaping damping matrix using the same direction vector constraint, switch the robotic arm to admittance control mode, and rely on the body torque sensing to perform impedance defense and passive safety slow stop.

[0189] The impedance safety defense module not only recalculates the spatial stiffness of the system but also performs matching adjustments to the system's motion damping characteristics using the same constraints. The module's computational unit continues to use the generated error direction unit vector to perform consistent outer product recombination calculations on the fundamental damping parameters within the underlying controller. By superimposing damping gains along the exact same error divergence direction, the computational unit calculates and generates a plastic damping matrix that strictly corresponds to the plastic stiffness matrix, used to suppress sudden velocity changes and high-frequency oscillations of the robotic arm in that direction.

[0190] After calculating the plastic stiffness matrix and plastic damping matrix, the impedance safety defense module generates a mode switching command and sends it to the underlying servo controller via the control system's communication bus. This mode switching command forces the underlying controller to change its operating mode from the original position control mode, which aims for precise spatial tracking, to an admittance control mode that provides flexible compliance based on force feedback.

[0191] Upon entering admittance control mode, the underlying controller begins to read and process in real time the actual sensing data from the underlying torque sensors installed on each joint of the robotic arm or the end effector. The impedance safety defense module injects the calculated plastic stiffness matrix and plastic damping matrix as new control parameters into the admittance control algorithm. Under the combined effect of torque feedback data and reshaped impedance parameters, the robotic arm exhibits basic flexible compliance when facing contact with the external environment, thus presenting a high stiffness and high damping state in the specific direction of residual divergence, thereby achieving passive safety slow stop and effectively avoiding equipment contact damage and abnormal displacement.

[0192] Specific application examples:

[0193] A specific application example of robotic arm control under welding dust obstruction and drastic changes in lighting conditions. In this application scenario, the front-end asynchronous data alignment module of the control system continuously acquires the bottom-level state of a six-degree-of-freedom robotic arm at a fixed high-frequency control cycle.

[0194] At a certain absolute timestamp At 5.250s, the system obtains the current absolute rotation angle, i.e., the current joint angle vector. rad, and the corresponding current joint angular velocity vector rad / s. The forward kinematics solving unit within the system utilizes the fundamental translational Jacobian matrix under this physical configuration. Perform kinematic mapping calculations to obtain the current actual linear velocity vector of the robotic arm's end effector. m / s. At the same time, the industrial camera transmitted a noisy image due to high concentration of dust. The control system accurately recorded the exposure midpoint time and obtained the corresponding visual prediction pose data through forward inference by a deep learning network.

[0195] Subsequently, the control system performs coupled disturbance feature extraction and error quantification. The coupled feature extraction module traces back historical states based on the aforementioned exposure timestamps and performs a differential comparison with the visual derivation results to calculate the aligned residual linear velocity vector. m / s. During this period, the calculated high-frequency Shannon entropy of the image is... The value increased to 7.85, indicating a relatively severe degree of visual degradation. The adaptive filtering module extracts the L2 norm scalar of the aligned residual linear velocity vector to obtain the residual linear velocity amplitude. =0.108 m / s. The program calculates and generates the dynamic observation noise covariance matrix by substituting preset parameters. Its matrix trace increases from a small value in the normal state to 0.65. The Kalman filter receives this inflated covariance parameter and actively reduces the trust path to erroneous visual guidance data.

[0196] When the residual linear velocity amplitude =0.108m / s exceeds the system's fixed safety constraint threshold of 0.100m / s, forcibly triggering the impedance safety defense logic. The defense module generates a unit vector in the error direction. The underlying controller extracts the reference stiffness matrix. N / m, and substitute the set stiffness penalty scalar. =5000N / m was used for calculation, combined with and its transpose vector Finally, the plastic stiffness matrix is ​​calculated and generated. .

[0197] See attached document Figure 8 The horizontal axis of the graph represents time (s), ranging from 0 to 10; the vertical axis represents the velocity deviation amplitude (m / s), ranging from 0 to 0.3. The black lines marked with solid triangles represent the velocity deviation evolution trajectory under traditional fixed-gain control, while the black lines marked with hollow squares represent the velocity deviation evolution trajectory after applying the adaptive control algorithm of this invention.

[0198] In-depth observation Figure 8As shown in the waveforms of the corresponding curves in the figure, during the normal, interference-free period of 0 to 3 seconds in the simulation, the velocity deviation amplitude of both curves remained stable at around 0.05 m / s. However, in the interval from 3 to 7 seconds when dense dust simulated blockage occurred, the traditional fixed-gain control curve marked with a solid triangle began to diverge upwards accompanied by high-frequency sawtooth oscillations. Its velocity deviation amplitude reached approximately 0.25 m / s around 4.2 seconds, and then peaked at approximately 0.26 m / s around 4.8 and 5.2 seconds, before finally falling back to around 0.05 m / s after 7 seconds. In contrast, the adaptive control curve of this invention, marked with a hollow square, effectively suppressed the transmission of erroneous information within the same interference interval of 3 to 7 seconds. The system's velocity deviation amplitude is stable, with only a small peak of about 0.08 m / s appearing at 4.8 seconds. For the rest of the time, it is locked near the safe baseline of 0.05 m / s and never exceeds 0.1 m / s in the complete 10-second cycle, which proves the strong anti-disturbance robustness brought about by the aforementioned formula calculation of the present invention.

[0199] Experimental verification and effect comparison:

[0200] To verify the actual physical protection effect of the impedance safety defense mechanism proposed in this invention (i.e., the plastic stiffness matrix and plastic damping matrix generated in the aforementioned sub-step S503) under extreme conditions such as system failure, a set of destructive comparative experiments with hard physical intervention was established. In the experimental setup, the end effector of the robotic arm was artificially induced to collide with a rigid metal obstacle equipped with a high-precision mechanical sensing test platform, and the force data was recorded in parallel.

[0201] See attached document Figure 9 The horizontal axis of the figure represents spatial displacement (m), ranging from 0 to 0.4; the vertical axis represents end-contact force (N), ranging from 0 to 60. The curves marked with solid circles record in detail the force evolution when physical interference occurs in the undefended position servo control mode; the curves marked with stars objectively show the force evolution after the admittance easing control of this invention is activated.

[0202] extract Figure 9Key data points reveal that during the free-space phase, from 0 to 0.2m, the contact force at the ends of both curves remains at 0N. At the instant the spatial displacement reaches 0.2m, the robotic arm's end effector first contacts a rigid obstacle. At this point, the unprotected position servo control curve, marked with solid circles, exhibits a steep and uncontrollable non-linear climb. The force on this curve is approximately 8N at 0.25m, increasing to approximately 18N at 0.3m, and to approximately 33N at 0.35m. Finally, at 0.4m, the contact force reaches 55N and shows a destructive upward trend, potentially leading to mechanical structural fracture.

[0203] Conversely, the admittance easing control curve of this invention, using a star-shaped marker, immediately exhibits physical yielding and flexible compliance along the collision negative feedback axis after the spatial displacement reaches 0.2m and initial physical contact occurs, based on the reshaped impedance parameters. The contact force of this curve not only does not experience a destructive increase, but instead generates only a small safety peak of approximately 7N at a displacement of about 0.23m. After crossing this peak, the force curve rapidly undergoes exponential unloading attenuation, and when the spatial displacement advances to 0.3m, the end-contact force has been strongly suppressed and returned to a position close to 0N. In the subsequent displacement range of 0.3m to 0.4m, the star-shaped marker curve consistently extends smoothly close to the horizontal axis (0N). This data comparison realistically reflects that this scheme successfully performs damage-free passive safety easing during the physical contact phase, confirming the absolute reliability of the invention's defense mechanism from a practical engineering perspective.

Claims

1. A vision-based deep learning adaptive control system for robotic arms operating in complex conditions, characterized in that, include: The asynchronous data alignment module records the absolute timestamp of the camera exposure midpoint and outputs visual prediction pose data, pushing the robot arm's current actual linear velocity vector into the first-in-first-out circular buffer. The coupled feature extraction module backtracks the historical actual linear velocity vector based on the absolute timestamp of the camera exposure midpoint, calculates the aligned residual linear velocity vector, and outputs the high-frequency Shannon entropy of the image. The adaptive filtering module calculates the dynamic observation noise covariance matrix based on the residual linear velocity amplitude generated by the high-frequency Shannon entropy of the image and the aligned residual linear velocity vector, and outputs the corrected pose control command to perform adaptive control on the robotic arm. The active scanning detection module outputs a microscopic scanning displacement vector trajectory command and a visual escape command to drive the robotic arm to adjust its motion trajectory when the trace of the dynamic observation noise covariance matrix exceeds a preset trace limit and the absolute value of the discrete-time first derivative of the high-frequency Shannon entropy of the image is less than a preset derivative threshold. The impedance safety defense module, when the residual linear velocity amplitude exceeds the preset safety constraint threshold, performs passive safety slow-stop on the robotic arm based on the aligned residual linear velocity vector combined with the admittance control mode. The coupled feature extraction module performs the calculation of the aligned residual linear velocity vector by backtracking the historical actual linear velocity vector based on the absolute timestamp of the camera exposure midpoint. Specifically, this includes: Using the absolute timestamp of the camera exposure midpoint as the retrieval condition, the system addresses the first-in-first-out circular buffer and backtracks to read the historical actual linear velocity vector that strictly corresponds to the absolute timestamp of the camera exposure midpoint. A time-different operation is performed on multiple continuously received visual prediction pose data to derive the visual derivation line velocity vector corresponding to the absolute timestamp of the camera exposure midpoint. Calculate the vector difference between the visual derivation line velocity vector and the historical actual line velocity vector to generate the alignment residual line velocity vector that eliminates the calculation misalignment error; The high-frequency Shannon entropy of the output image executed by the coupled feature extraction module specifically includes: Acquire a single frame image captured in parallel, and perform a two-dimensional fast Fourier transform on the spatial domain two-dimensional pixel matrix of the single frame image to transform the single frame image into a two-dimensional frequency domain space. Extract high-frequency mask region data above the set cutoff frequency in the two-dimensional frequency domain space to construct the high-frequency power spectrum of the current image frame; Normalize the high-frequency power spectral density data to generate a normalized high-frequency power spectral density; The normalized high-frequency power spectral density is combined with a logarithmic function and a summation operation to calculate and output the high-frequency Shannon entropy of the image that quantifies the external optical degradation state. The adaptive filtering module performs the calculation of the dynamic observation noise covariance matrix based on the residual linear velocity amplitude generated by the high-frequency Shannon entropy of the image and the aligned residual linear velocity vector. Specifically, this includes: Extract the three-dimensional coordinate components of the aligned residual linear velocity vector and calculate the square root of the sum of squares of each coordinate component to convert it into a scalar value, thereby generating the residual linear velocity amplitude; A modulation function containing a bidirectional exponential decay term is executed, the first term of which is used to calculate the covariance inflation factor caused by environmental optical degradation based on the high-frequency Shannon entropy of the image; The second term of the modulation function is based on the extracted residual linear velocity amplitude, and the covariance inflation factor caused by the deviation between the visually derived state and the actual motion state of the body is calculated using the inverse exponential decay function. The penalty gain of the first and second terms is superimposed to generate the dynamic observation noise covariance matrix based on the baseline observation noise covariance matrix.

2. The vision-deep learning adaptive control system for robotic arms under complex working conditions according to claim 1, characterized in that, The asynchronous data alignment module records the absolute timestamp of the camera exposure midpoint and outputs visual prediction pose data. Specifically, it pushes the robot arm's current actual linear velocity vector into the first-in-first-out circular buffer. Establish an absolute time axis and continuously read the state data of the underlying robotic arm's motion joints according to a fixed high-frequency control cycle; Calculate the current actual linear velocity vector of the robotic arm's end effector based on the stated state data and the forward kinematics model; The current actual linear velocity vector is combined with the corresponding high-frequency timestamp data to form a motion state frame, and the motion state frame is pushed into the first-in-first-out circular buffer. The visual prediction pose data is output by processing asynchronously input image frames from an industrial camera using a deep learning network model, and the equivalent center time of the image frame capturing external optical information is recorded as the absolute timestamp of the camera exposure midpoint.

3. The vision-based deep learning adaptive control system for robotic arms under complex working conditions according to claim 1, characterized in that, The adaptive filtering module executes the corrected pose control command to perform adaptive control on the robotic arm, specifically including: The generated dynamic observation noise covariance matrix is ​​input into the internal extended Kalman filter as the observation noise parameter; The current Kalman gain matrix is ​​calculated using the Kalman gain equation, which automatically reduces the weight of the dependence on the current visual prediction pose data. By combining the reduced-weight visual observation data with the state prediction data based on the kinematic model, the pose state of the robotic arm end effector is smoothly estimated and the state recursion is calculated to generate the corrected pose control command. The corrected pose control command is sent to the underlying servo driver to control the movement trajectory of the robotic arm.

4. The vision-deep learning adaptive control system for robotic arms under complex working conditions according to claim 1, characterized in that, The active scanning detection module performs the following actions when the trace of the dynamic observation noise covariance matrix exceeds a preset trace limit and the absolute value of the discrete-time first derivative of the high-frequency Shannon entropy of the image is less than a preset derivative threshold: The trace of the dynamic observation noise covariance matrix is ​​obtained by reading the dynamic observation noise covariance matrix in real time and calculating the sum of the main diagonal elements of the dynamic observation noise covariance matrix. The high-frequency Shannon entropy sequence of the image for multiple consecutive control cycles is obtained, and the discrete-time first derivative is calculated using a numerical differentiation algorithm. When it is determined that the trace of the dynamic observation noise covariance matrix exceeds the preset trace limit and the absolute value of the discrete-time first derivative of the high-frequency Shannon entropy of the image is less than the preset derivative threshold within multiple consecutive control cycles, a control cutoff command is issued to suspend the current position servo control.

5. The vision-deep learning adaptive control system for robotic arms under complex working conditions according to claim 4, characterized in that, After pausing the current position servo control, the scanning active detection module is also used for: The microscopic scanning displacement vector trajectory command containing orthogonal periodic signals with different frequencies and amplitudes is applied to the underlying controller to drive the end effector of the robotic arm to perform spatial periodic oscillation motion with a preset amplitude. During the oscillating motion, the theoretical instantaneous velocity vector calculated is continuously monitored. When the magnitude of the three-dimensional coordinate component drops to zero or below the preset velocity threshold, it is determined that the instantaneous zero-crossing point in three-dimensional space has been reached, and a hard trigger command is issued through the hardware pin to force the industrial camera to conduct a short exposure to acquire the image. By comparing the high-frequency Shannon entropy of the image acquired through continuous discrete zero-speed point exposures, the visual escape command is output according to the negative gradient direction in the discrete space to reconstruct the subsequent control trajectory.

6. The vision-deep learning adaptive control system for robotic arms under complex working conditions according to claim 1, characterized in that, The impedance safety defense module performs passive safety easing of the robotic arm when the residual linear velocity amplitude exceeds a preset safety constraint threshold, based on the aligned residual linear velocity vector and admittance control mode. Specifically, this includes: The residual linear velocity amplitude is monitored in real time. When it is determined that the residual linear velocity amplitude exceeds the preset safety constraint threshold, a global safety back-off response mechanism is triggered and the coordinate data output of the adaptive filtering module is stopped. Obtain the alignment residual linear velocity vector and perform a normalization operation on the alignment residual linear velocity vector to generate an error direction unit vector characterizing the error divergence direction; The outer product of the error direction unit vector and the transpose of the error direction unit vector is calculated to construct the direction gain matrix. The direction gain matrix is ​​multiplied by the set stiffness penalty scalar and superimposed on the reference stiffness matrix to generate the plastic stiffness matrix. By performing the same outer product recombination operation on the basic damping parameters using the error direction unit vector, a plastic damping matrix that strictly corresponds to the plastic stiffness matrix is ​​generated.

7. The vision-deep learning adaptive control system for robotic arms under complex working conditions according to claim 6, characterized in that, After generating the plastic stiffness matrix and the plastic damping matrix, the impedance safety defense module is also used for: The generation mode switching command is sent to the underlying controller to switch the operating mode from the position control mode to the admittance control mode; In the admittance control mode, the actual sensing data of the underlying torque sensors installed on each motion joint or end effector of the robotic arm are read in real time. The plastic stiffness matrix and plastic damping matrix are injected as new control parameters into the admittance control algorithm; Under the combined effect of torque feedback data and reshaped impedance parameters, the robotic arm exhibits flexible compliance when facing the external environment, while displaying high stiffness and high damping in the specific direction of residual divergence, thereby achieving passive safe slow stop.

Images