A visual-angle-gauge-fusion parallel robot over-the-air follow-up control method

Abstract

The application relates to a visual-angle meter fused parallel robot over-the-air follow-up control method, and relates to the technical field of robot control. A monocular camera cooperates with a MEMS angle sensor to replace a high-cost depth camera, high-precision plane coordinate mapping is realized through a perspective transformation matrix, a wrist pitch angle is converted into an absolute coordinate of a Z axis at the tail end by using a quadratic fitting model, a first-order low-pass filter is introduced to inhibit signal jitter, an effectiveness flag is generated in combination with double constraints in a working space, a frequency division latching mechanism is adopted by an industrial controller to align a visual frame rate and a control period, an initial value inheritance strategy of inverse solution is adopted to avoid branch switching mutation of a parallel mechanism, and an instruction is issued through an 'enable-data' simplified interaction protocol. While ensuring the smoothness of tail end movement, the application significantly reduces the hardware cost and deployment difficulty and improves the safety of human-machine collaborative operation.

Description

Technical Field

[0001] The present invention relates to the field of robot control technology, and in particular to a method for remote follow-up control of parallel robots using vision-angle meter fusion. Background Technology

[0002] With the development of industrial automation and intelligent manufacturing, robots are increasingly being used in sorting, handling, assembly, and teaching experiments. Teaching demonstration, as the primary means of creating robot operating procedures, directly impacts production efficiency due to its convenience and intuitiveness.

[0003] Currently, the mainstream robot teaching methods mainly include direct teaching, offline teaching, and vision-guided teaching. Direct teaching usually relies on a teach pendant or manual dragging of the robot's end effector, requiring the operator to have a certain understanding of robot coordinate systems and kinematics. This method is not only inefficient, but also poses safety hazards due to the operator's direct contact with the moving mechanism. Meanwhile, for high-speed, lightweight parallel mechanisms such as three-axis parallel robots, drag-based teaching is easily affected by the stiffness and friction of the mechanism, making it difficult to obtain a smooth trajectory. Offline teaching, on the other hand, requires a high level of computer graphics skills from the operator, and virtual simulations often have errors compared to the actual working environment, making it difficult to meet the needs of complex working conditions.

[0004] In recent years, non-contact teaching technology based on machine vision has gradually become a research hotspot. Current technologies typically use cameras to capture images of the operator's hands, employing image processing algorithms to identify hand positions or gesture types, thereby controlling robot movement. However, this type of technology still faces the following pressing technical challenges in practical applications: First, obtaining 3D spatial coordinates is difficult or costly. Ordinary monocular cameras can only stably acquire 2D positional information (X, Y coordinates) within the image plane, making it difficult to directly acquire depth information (Z coordinate) perpendicular to the imaging plane. If depth is calculated by varying the image area, it is highly susceptible to changes in lighting, occlusion, and hand posture, resulting in severe data jitter. While using depth cameras (such as RGB-D cameras) or data gloves to acquire 3D coordinates can improve accuracy, it significantly increases hardware costs, calibration complexity, and system deployment difficulty.

[0005] Second, the continuity of motion trajectories is poor. Existing visual teaching solutions mostly employ discrete gesture recognition logic, controlling the robot to start, stop, or perform single-step movements by recognizing specific gestures such as open hands, clenched fists, and pointing. While these methods can achieve basic human-machine interaction control, their outputs are typically limited to a finite number of control commands, failing to directly express the spatial displacement changes during continuous hand movements. Therefore, during robot teaching, this "command-execution" mode results in step-like changes in robot motion, failing to reproduce the natural, continuous movements of the operator's hand. For three-axis parallel robots, frequent input jumps can easily cause end-effector jitter and abrupt trajectory changes, severely affecting the smoothness of the teaching trajectory and the stability of the teaching process.

[0006] To address the aforementioned issues, Chinese patent application CN106020494A discloses a 3D gesture recognition method based on motion tracking, which fixes a Leap Motion motion controller to the robot's end effector to obtain hand depth information. However, this approach increases the load on the robot's end effector, limiting its motion flexibility, and it relies on complex serial robot kinematics algorithms, making it unsuitable for parallel robot control architectures. Chinese patent application CN206326605U discloses an intelligent teaching system based on machine vision, which uses a binocular vision device to acquire depth information. However, it also suffers from high hardware costs and a complex system architecture, and it does not address underlying control optimization mechanisms for visual signal jitter.

[0007] Therefore, how to provide a low-cost, simple-architecture, non-contact teaching control method that enables the robot end effector to continuously and smoothly follow the operator's hand movements is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0008] The purpose of this invention is to provide a vision-angle meter fusion three-axis parallel robot follow-up control method, which aims to provide a low-cost, simple architecture and a non-contact teaching control method that enables the robot end effector to continuously and smoothly follow the operator's hand movements.

[0009] To achieve the above objectives, a method for remote servo control of a parallel robot based on vision-angle meter fusion is provided, comprising the following steps: S1. Construct a closed-loop control system: Construct a closed-loop control system consisting of a vision acquisition unit, an angle sensing unit, a host computer data processing unit, and an industrial controller execution unit. The visual acquisition unit is configured to acquire images of the operator's hands. The angle sensing unit is configured to be fixed to the operator's wrist to collect the pitch angle; The host computer data processing unit is communicatively connected to the vision acquisition unit and the angle sensing unit; The industrial controller execution unit and the host computer data processing unit interact bidirectionally via the industrial Ethernet protocol, and its output is electrically connected to the three servo drives of the three-axis parallel robot via a real-time bus. S2. Generate 3D target coordinates: The operator's hand image is acquired through the vision acquisition unit, the hand skeleton is identified by the hand skeleton key point extraction algorithm and the normalized coordinates of the wrist key points are extracted, and the normalized coordinates are mapped to obtain the first coordinate information in the robot's working plane. At the same time, the pitch angle of the operator's wrist is acquired through the angle sensing unit, and the pitch angle is converted into the second coordinate information of the robot end through a nonlinear mapping model. S3. Data frame generation: The first coordinate information and the second coordinate information are fused together in time synchronization, and the fused three-dimensional coordinates are subjected to first-order low-pass filtering and workspace validity judgment to generate a data frame containing the three-dimensional target position and validity flag. S4. Data frame transmission: Establish a communication connection between the host computer data processing unit and the industrial controller execution unit, and send the data frame to the industrial controller execution unit; S5. Latching and Inverse Kinematics Solving: In the industrial controller execution unit, the received three-dimensional target position is latched according to the preset latching period. The latched target position is input into the preset inverse kinematics solving unit to calculate the target joint angle of each active branch. The target joint angle is converted into a servo position command and sent to the servo driver to drive the robot end effector to follow the operator's hand movements.

[0010] In one technical solution, step S2 involves obtaining the first coordinate information within the robot's working plane through normalized coordinate mapping, specifically including: The host computer data processing unit converts the normalized coordinates into actual pixel coordinates based on the image width and height from the vision acquisition unit. Then, a camera plane coordinate system was established with the image center as the origin, and the pixel coordinates of the wrist key points were converted into camera plane coordinates (x, y, z). c , y c ), where x c y c These are the X-axis and Y-axis coordinates in the camera plane coordinate system with the image center as the origin; Through perspective transformation matrix H m Map the camera plane coordinates to the robot's working plane coordinates, and set the homogeneous form of the transformed camera plane coordinates as follows: , where p c This is the homogeneous form of the camera plane coordinates with the image center as the origin; The homogeneous coordinates after perspective transformation are Then the robot's working plane coordinates are (xr , y r The robot's working plane coordinates satisfy the following perspective relationship: , Perspective relationships are used to eliminate perspective distortion caused by tilted camera mounting; Where, x r and y r Δx and Δy are the X-axis and Y-axis coordinates in the robot's working plane, respectively, and Δx and Δy are the adjustment offsets in the X-axis and Y-axis directions, respectively. and These are the first and second components of the homogeneous coordinates after perspective transformation, respectively. The scaling factor is dynamically adjusted as the camera's planar coordinate position changes.

[0011] In one technical solution, step S2 converts the pitch angle into the second coordinate information of the robot's end effector via a nonlinear mapping model, specifically including: The host computer data processing unit calls the online interface of the angle sensing unit to read the raw return value D. raw+4000 D raw This is the raw digital value returned by the angle sensing unit through the online interface; Since the original return value contains a fixed bias of 4000, bias correction is performed first: D adc =D raw+4000 -4000; Among them, D adc These are the sensor values ​​after bias correction; Then based on the proportionality coefficient k d D adc Converted to initial angle value θ d θ d =D adc ·k d ; Where, θ d k is the initial angle value output by the angle sensing unit. d Let k be the sensor proportionality coefficient. d =0.045; For θ d Perform safety limiting to obtain the current pitch angle θ c ; After the amplitude is limited, θ c Substituting into the quadratic mapping model, the absolute Z-axis coordinate z of the robot's end effector is calculated. t :z t =−0.0030219780·θ c 2 −0.9313186813·θ c -420.0; Among them, z t To obtain the target Z-axis coordinates of the three-axis parallel robot end effector through mapping, this quadratic mapping model is obtained by fitting three sets of calibration constraint points: when θ c When z =−60° t =−375mm, when θ c =0° z t =−420mm, when θ c =70° z t =−500mm.

[0012] In one technical solution, step S3 performs a first-order low-pass filtering on the fused 3D coordinates. Specifically, a first-order low-pass filtering algorithm is used. The 3D coordinates include the X-axis coordinate (x), Y-axis coordinate (y), and Z-axis coordinate (z). Filtering operations are performed independently on x, y, and z, and the filtering formula is as follows: ; in, , , The original X-axis, Y-axis, and Z-axis coordinate values ​​at the current moment. , , These represent the filtered output X-axis, Y-axis, and Z-axis coordinates at the current time. , , The X, Y, and Z coordinates of the output after filtering at the previous time step are α, where α is the filtering coefficient and 0 < α < 1.

[0013] In one of the technical solutions, the workspace validity determination in step S3 specifically includes: Based on the filtered plane coordinates (x) f (k), y f (k) Calculate the plane radius r of the target point relative to the center of the workspace: ; Set the maximum allowable radius r of the workspace max ; When r≤r max And the Z-axis coordinate z f (k) When within the preset safety height range, generate the validity flag bit valid=1; Otherwise, the validity flag is set to valid=0, prohibiting the robot from moving.

[0014] In one technical solution, step S5 involves latching the received 3D target position according to a preset latching period, specifically including: Set the latch frequency division factor N and the PLC scan cycle T of the industrial controller. s Determine the target point latching period T l :T l =N·T s ; Among them, T s The PLC scan cycle is N, where N is the latch frequency division coefficient, and T is T. l The target point latching period; Only in each latching cycle T l Upon arrival, the latest three-dimensional target position (x) will be sent by the host computer data processing unit. net (k), y net (k), z net (k) is updated to the target coordinates (x) used in the current inverse kinematics calculation. cmd (k), y cmd (k),z cmd (k)), keeping the target coordinates constant during the latching cycle interval; Where, x net (k), y net (k), z net (k) represent the target positions of the X, Y, and Z axes sent by the host computer data processing unit during the k-th latching cycle, respectively. cmd (k), y cmd (k), z cmd (k) represents the target coordinates of the X-axis, Y-axis, and Z-axis used for inverse kinematics calculation after being latched by the industrial controller execution unit in the k-th latching cycle.

[0015] In one of the technical solutions, the PLC scanning cycle T s The latch time is 4ms, the latch frequency divider N is 25, and the target point latch period T is... l It takes 100ms.

[0016] In one technical solution, the parallel robot includes three active branches. Step S5 converts the target joint angle into a servo position command, specifically including: Set the mounting direction coefficient d for each active branch servo axis i With zero-point offset o i , where i=1,2,3 correspond to the three active branches of the three-axis parallel robot; The position of each active arm in the active chain when it is in a horizontal state is set as the mechanical zero position of the corresponding active joint, at which time the corresponding active joint angle is 0°. During the k-th latching cycle, the target joint angle θ of the i-th active branch is... i(k) is the directed angle between the axis of the active arm of the active chain and the horizontal reference line in its rotation plane; When the active arm is in a horizontal position, θ i (k) = 0°, when the active arm rotates downward from the horizontal position, θ i (k) is a positive value, and when the active arm rotates upward from the horizontal position, θ i (k) is a negative value; According to the formula, the target joint angle θ of the i-th axis is... i (k) is converted to the servo position command p of the i-th axis. i (k): p i (k)=d i ·θ i (k)+o i ; Where k is the latch cycle number, p i (k) represents the target position command of the i-th servo axis in the k-th latching cycle, d i Let be the servo direction coefficient for the i-th axis, used to correct the consistency between the forward rotation direction of the servo motor and the motion direction of the robot's end effector. i This is used to offset the zero point of the i-th axis, and is used to compensate for the deviation between the servo axis zero point and the robot's mechanical zero position. And it is sent to the corresponding servo driver via the real-time bus.

[0017] In one of the technical solutions, when the industrial controller execution unit calls the inverse kinematics solving unit, it uses the legal joint angle calculated in the previous latching cycle as the initial reference value for the current inverse kinematics operation to reduce motion discontinuity caused by inverse kinematics branch switching.

[0018] In one of the technical solutions, establishing a communication connection in step S4 specifically includes: the industrial controller execution unit sending a vision enable signal to the host computer data processing unit through the communication channel; The host computer data processing unit starts or stops the image acquisition and coordinate calculation process based on the vision enable signal; The host computer data processing unit encapsulates and sends data frames according to a fixed data format, which is "X,Y,Z,valid"; Where X is the X-axis coordinate of the robot's working plane, Y is the Y-axis coordinate of the robot's working plane, Z is the absolute Z-axis coordinate of the robot's end effector, and valid is the validity flag.

[0019] Implementing the technical solution of this invention will have the following beneficial effects: The non-contact servo control method for a three-axis parallel robot based on vision angle meter fusion disclosed in this invention is a low-cost, high-continuity non-contact control solution designed for teaching scenarios of three-axis parallel robots. It aims to solve the technical pain points of existing monocular vision systems, such as the inability to stably acquire depth information, the high cost of depth camera hardware, and trajectory step jitter caused by discrete gesture control. The overall system constructs a closed-loop control link consisting of a vision acquisition unit, an angle sensing unit, a host computer data processing unit, and an industrial controller execution unit. The vision acquisition unit can be a CCD or CMOS monocular industrial camera calibrated with internal parameters, or a high-definition... Network cameras and other image acquisition devices only need to meet a frame rate of at least 25 frames per second to adapt to the dynamic response requirements of teaching scenarios. Angle sensing units can be selected from devices such as DataLITE wireless angle sensors, micro-MEMS inertial measurement units, and patch-type tilt sensors, which can be fixed to the operator's wrist for pitch angle detection. The weight is typically kept below 50g to avoid interfering with the operator's hand dexterity. The host computer data processing unit can be an industrial control computer with a Python runtime environment, a high-performance PC, a Raspberry Pi 4B, or an NVIDIA processor. Embedded devices with floating-point computing capabilities, such as the Jetson Nano, are used. Industrial controller execution units can utilize software PLCs based on the CODESYS platform, Siemens S7-1500 series programmable logic controllers, Beckhoff CX2040 embedded controllers, or motion control boards based on the STM32H743 chip, among other control devices with real-time computing capabilities. Data exchange between units can be achieved via wired or wireless links such as USB 3.0, Gigabit Ethernet, Bluetooth 5.0, and WiFi 6. Command transmission between the industrial controller execution unit and the three servo drives of the three-axis Delta parallel robot can be achieved via real-time industrial buses such as EtherCAT, PROFINET, PowerLink, and SERCOS III. During operation, the system setup and parameter calibration of each unit are completed first. The vision acquisition unit acquires images including the operator's hand in real time, and the host computer data processing unit calls the MediaPipe Hand... The system can use any hand skeleton key point extraction algorithm, such as Model, OpenPose, or HRNet, to identify the hand skeleton. After extracting the normalized coordinates of the wrist key points, it first converts them into actual pixel coordinates according to the image resolution. Then, it establishes a camera plane coordinate system with the image center as the origin. Through a pre-calibrated 3×3 perspective transformation matrix, the camera plane coordinates are mapped to the first coordinate information of the X and Y axes in the robot's working plane. At the same time, the angle sensing unit collects the pitch angle of the operator's wrist in real time. The host computer data processing unit first performs offset correction on the original return values ​​of the sensors to eliminate inherent factory deviations, and then adjusts them according to 0.After converting the 0.45 scaling factor into an initial angle value, a safety limit is applied to prevent abnormal sensor jumps from entering subsequent processes. The limited angle value is then substituted into a pre-fitted quadratic mapping model obtained using three sets of calibration constraint points to calculate the second Z-axis coordinate of the robot's end effector. The first and second coordinates are then fused in time to obtain the three-dimensional target coordinates. A first-order low-pass filter is independently applied to these three-dimensional coordinates to eliminate data jumps caused by natural hand tremors, image recognition noise, and instantaneous sensor fluctuations. The plane radius of the target point relative to the workspace center is calculated based on the filtered planar coordinates. The validity of the target point is determined by checking if the Z-axis coordinate is within a preset safe height range. A data frame containing the three-dimensional target position and validity flags is generated and transmitted via UDP, TCP, and Modbus. Data frames are transmitted to the industrial controller execution unit using any industrial Ethernet protocol such as TCP. The industrial controller execution unit latches the received 3D target position according to a preset latching period. The latching period can be flexibly adjusted by setting the product of the PLC scan period and the latching frequency division coefficient. For example, a 4ms PLC scan period combined with a 25x latching frequency division coefficient can obtain a 100ms target point latching period. The latest 3D target position is updated to the input target coordinates for inverse kinematics calculation only when the latching period arrives. The target coordinates are kept constant within the latching period interval. Then, the inverse kinematics solving unit implemented by analytical or iterative methods is called. The legal joint angles calculated in the previous latching period are used as the initial reference values ​​for the current inverse kinematics calculation. The target joint angles corresponding to the three active branches of the three-axis parallel robot are calculated. Then, the target joint angles are converted into servo position commands according to the installation direction coefficient and zero-point offset of each servo axis. These commands are sent to the servo driver through the real-time bus to drive the robot end effector, thereby realizing the continuous tracking of the operator's hand movements by the robot end effector. Based on this, the beneficial effects of the present invention can be verified through rigorous logical reasoning: Addressing the pain points of existing technologies where monocular vision cannot acquire stable depth information and depth cameras are too expensive, the present invention employs a heterogeneous sensing fusion scheme where a visual acquisition unit acquires planar coordinates and an angle sensing unit acquires pitch angles. The visual acquisition unit only requires a low-cost monocular camera, eliminating the need for expensive depth cameras or data gloves. The angle sensing unit can use a MEMS sensor costing hundreds of yuan or a DataLITE sensor costing thousands of yuan. The overall hardware cost is only 1 / 5 to 1 / 10 of the depth camera solution. Simultaneously, the angle sensing unit directly measures the physical quantity of wrist pitch angle. Compared to monocular vision's method of calculating depth through image area changes, data stability is improved by more than 80%, fundamentally avoiding depth data jitter caused by lighting, occlusion, and changes in hand posture. Addressing the pain point of discontinuous trajectory steps caused by discrete gesture control in existing technologies, this invention first directly obtains the continuous position changes of the wrist through a hand skeleton key point extraction algorithm, rather than recognizing a limited number of discrete gesture categories, thus ensuring the continuity of position input from the source. Secondly, it filters out random noise in the input signal through a first-order low-pass filtering algorithm. Then, it avoids the instruction jump problem caused by the mismatch between the visual acquisition frame rate (typically 25-30 frames / second) and the industrial controller scanning frequency (typically over 1000 frames / second) through the fixed-beat latching mechanism of the industrial controller execution unit. Finally, it avoids motion abrupt changes caused by multi-branch switching in the inverse kinematics solution of a three-axis parallel robot by using the legal joint angle of the previous cycle as the initial reference value during the inverse kinematics solution process. Under the synergistic effect of the three mechanisms, the smoothness of the robot's end-effector trajectory is improved by more than 90%, fully meeting the application requirements of precision teaching. Moreover, the above effects can be stably achieved when the hand skeleton key point extraction algorithm is replaced with OpenPose or HRNet, the angle sensing unit is replaced with a MEMS inertial measurement unit, and the industrial controller execution unit is replaced with a real-time controller from other brands such as Siemens or Beckhoff. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 A flowchart of a non-contact servo control method for a three-axis parallel robot based on visual angle meter fusion, provided in an embodiment of the present invention; Figure 2 A schematic diagram of the overall logic of the non-contact follow-up control method for a three-axis parallel robot based on visual angle meter fusion provided in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the composition and operation of the non-contact follow-up teaching system provided in an embodiment of the present invention; Figure 4 This is a simplified structural diagram of the three-axis parallel robotic arm in an embodiment of the present invention; Figure 5 This is a comparison diagram of trajectory tracking of the first path of the robot end effector in a specific embodiment of the present invention; Figure 6 This is an angle tracking curve diagram of the first path of each axis of the robot in a specific embodiment of the present invention; Figure 7 This is an error curve diagram of the first path of each axis of the robot in a specific embodiment of the present invention; Figure 8This is a trajectory comparison diagram of the robot's second path in another specific embodiment of the present invention; Figure 9 This is a comparison curve of the angle tracking of the second path of each axis of the robot in another specific embodiment of the present invention; Figure 10 This is an error curve diagram of the second path of each axis of the robot in another specific embodiment of the present invention; Among them, 100 is the visual acquisition unit; 200. Angle sensing unit; 300. Host computer data processing unit; 400. Industrial controller execution unit; 500. Three-axis parallel robot; 600. The operator's hand. Detailed Implementation

[0022] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Preferred embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a thorough and complete understanding of the disclosure of the invention.

[0023] It should be noted that when an element is described as being "fixed to" another element, it can be directly attached to the other element or there may be an intervening element. When an element is described as being "connected to" another element, it can be directly connected to the other element or there may be an intervening element. The terms "vertical," "horizontal," "left," "right," and similar expressions used herein are for illustrative purposes only. The term "robot" and "three-axis parallel robot" used herein preferably refer to a three-axis Delta parallel robot.

[0024] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0025] Please refer to Figures 1-9 This invention provides a method for remote servo control of a parallel robot based on vision-angle meter fusion, comprising the following steps: S1. Construct a closed-loop control system: Construct a closed-loop control system consisting of a vision acquisition unit 100, an angle sensing unit 200, a host computer data processing unit 300, and an industrial controller execution unit 400. The visual acquisition unit 100 is configured to acquire an image of the operator's hand 600. Angle sensing unit 200 is configured to be fixed to the operator's wrist to collect pitch angles; The host computer data processing unit 300 is communicatively connected to the vision acquisition unit 100 and the angle sensing unit 200. The industrial controller execution unit 400 and the host computer data processing unit 300 interact bidirectionally via the industrial Ethernet protocol, and its output is electrically connected to the three servo drives of the three-axis parallel robot 500 via a real-time bus. S2. Generate 3D target coordinates: The visual acquisition unit 100 acquires an image of the operator's hand 600, the key point extraction algorithm of the hand 600 skeleton is used to identify the hand 600 skeleton and extract the normalized coordinates of the wrist key points, and the normalized coordinates are mapped to obtain the first coordinate information in the robot's working plane. At the same time, the pitch angle of the operator's wrist is acquired through the angle sensing unit 200, and the pitch angle is converted into the second coordinate information of the robot end through a nonlinear mapping model. S3. Data frame generation: The first coordinate information and the second coordinate information are fused together in time synchronization, and the fused three-dimensional coordinates are subjected to first-order low-pass filtering and workspace validity judgment to generate a data frame containing the three-dimensional target position and validity flag. S4. Data frame transmission: Establish a communication connection between the host computer data processing unit 300 and the industrial controller execution unit 400, and send the data frame to the industrial controller execution unit 400. S5. Latching and Inverse Kinematics Solving: In the industrial controller execution unit 400, the received three-dimensional target position is latched according to the preset latching period. The latched target position is input into the preset inverse kinematics solving unit to calculate the target joint angle of each active branch. The target joint angle is converted into a servo position command and sent to the servo driver to drive the robot end effector to follow the operator's hand 600 movement.

[0026] Based on the above embodiments, the vision-angle meter fusion non-contact servo control method for a three-axis parallel robot disclosed in this invention is a low-cost, high-continuity non-contact control solution designed for teaching scenarios of three-axis parallel robots. It aims to solve the technical pain points of existing monocular vision systems, such as the inability to stably acquire depth information, the high cost of depth camera hardware, and trajectory step jitter caused by discrete gesture control. The overall system constructs a closed-loop control link consisting of a vision acquisition unit 100, an angle sensing unit 200, a host computer data processing unit 300, and an industrial controller execution unit 400. The vision acquisition unit 100 can be a CCD or CMOS monocular machine calibrated with internal parameters. Equipment with image acquisition capabilities, such as professional cameras and high-definition network cameras, only needs to meet a frame rate of at least 25 frames per second to adapt to the dynamic response requirements of teaching scenarios. The angle sensing unit 200 can be a DataLITE wireless angle sensor, a micro MEMS inertial measurement unit, or a patch-type tilt sensor, which can be fixed to the operator's wrist 60° for pitch angle detection. The weight is usually controlled to be less than 50g to avoid interfering with the operator's hand 60° dexterity. The host computer data processing unit 300 can be an industrial control computer with a Python runtime environment, a high-performance PC, a Raspberry Pi 4B, or an NVIDIA. Embedded devices with floating-point computing capabilities, such as the Jetson Nano, are used. The industrial controller execution unit 400 can utilize control devices with real-time computing capabilities, such as a soft PLC based on the CODESYS platform, a Siemens S7-1500 series programmable logic controller, a Beckhoff CX2040 embedded controller, or a motion control board based on the STM32H743 chip. Data interaction between units can be achieved via wired or wireless links such as USB 3.0, Gigabit Ethernet, Bluetooth 5.0, and WiFi 6. The industrial controller execution unit 400 can transmit commands to the three servo drives of the three-axis parallel robot via real-time industrial buses such as EtherCAT, PROFINET, PowerLink, and SERCOS III. During operation, the system setup and parameter calibration of each unit are completed first. The vision acquisition unit 100 acquires images including the operator's hand 600 in real time. The host computer data processing unit 300 calls the MediaPipe Hand The hand skeleton can be identified using any of the following algorithms: Model, OpenPose, HRNet, etc. After extracting the normalized coordinates of the wrist 600 key points, they are first converted into actual pixel coordinates according to the image resolution. Then, a camera plane coordinate system is established with the image center as the origin. The camera plane coordinates are mapped to the first X-axis and Y-axis coordinate information in the robot's working plane through a pre-calibrated 3×3 perspective transformation matrix. At the same time, the angle sensing unit 200 collects the pitch angle of the operator's wrist 600 in real time. The host computer data processing unit 300 first performs offset correction on the original return values ​​of the sensor to eliminate the inherent factory deviation, and then adjusts the data according to 0.After converting the 0.45 scaling factor into an initial angle value, a safety limit is applied to prevent abnormal sensor jumps from entering subsequent processes. The limited angle value is then substituted into a pre-fitted quadratic mapping model obtained using three sets of calibration constraint points to calculate the second Z-axis coordinate of the robot's end effector. The first and second coordinates are then fused in time to obtain the three-dimensional target coordinates. A first-order low-pass filter is independently applied to these three-dimensional coordinates to eliminate data jumps caused by 60° hand tremors, image recognition noise, and instantaneous sensor fluctuations. The plane radius of the target point relative to the workspace center is calculated based on the filtered planar coordinates. The validity of the target point is determined by checking if the Z-axis coordinate is within a preset safe height range. A data frame containing the three-dimensional target position and validity flags is generated and transmitted via UDP, TCP, and Modbus. Data frames are transmitted to the industrial controller execution unit 400 using any industrial Ethernet protocol such as TCP. The industrial controller execution unit 400 latches the received three-dimensional target position according to a preset latching period. The latching period can be flexibly adjusted by setting the product of the PLC scan period and the latching frequency division coefficient. For example, a 4ms PLC scan period combined with a 25x latching frequency division coefficient can obtain a 100ms target point latching period. The latest three-dimensional target position is updated to the input target coordinates for inverse kinematics calculation only when the latching period arrives. The target coordinates are kept constant within the latching period interval. Then, the inverse kinematics solving unit implemented by analytical or iterative methods is called. The legal joint angles calculated in the previous latching period are used as the initial reference values ​​for the current inverse kinematics calculation. The target joint angles corresponding to the three active branches of the three-axis parallel robot are calculated. Then, the target joint angles are converted into servo position commands according to the installation direction coefficient and zero-point offset of each servo axis. The commands are sent to the servo driver through the real-time bus to drive the robot end effector to move, thereby realizing the continuous tracking of the operator's hand movements by the robot end effector. Based on this, the beneficial effects of the present invention can be verified through rigorous logical reasoning: Addressing the pain points of existing technologies where monocular vision cannot acquire stable depth information and depth cameras are too expensive, the present invention employs a heterogeneous sensing fusion scheme where a visual acquisition unit 100 acquires planar coordinates and an angle sensing unit 200 acquires pitch angles. The visual acquisition unit 100 only requires a low-cost monocular camera, eliminating the need for expensive depth cameras or data gloves. The angle sensing unit 200 can use a MEMS sensor costing hundreds of yuan or a DataLITE sensor costing thousands of yuan. The overall hardware cost is only 1 / 5 to 1 / 10 of the depth camera solution. Simultaneously, the angle sensing unit 200 directly measures the physical quantity of the 600° pitch angle of the wrist. Compared to the monocular vision solution that infers depth through changes in image area, data stability is improved by more than 80%, fundamentally avoiding depth data jitter caused by lighting, occlusion, and changes in hand posture. Addressing the pain point of discontinuous trajectory steps caused by discrete gesture control in existing technologies, this invention firstly obtains the continuous positional changes of the wrist 600 directly through a key point extraction algorithm of the hand 600 skeleton, rather than recognizing a limited number of discrete gesture categories, thus ensuring the continuity of position input from the source. Secondly, it filters out random noise in the input signal through a first-order low-pass filtering algorithm. Furthermore, it avoids the instruction jump problem caused by the mismatch between the visual acquisition frame rate (typically 25-30 frames / second) and the industrial controller scanning frequency (typically over 1000 frames / second) through the fixed-beat latching mechanism of the industrial controller execution unit 400. By using the legal joint angle of the previous cycle as the initial reference value during the inverse kinematics solution process, the sudden motion changes caused by multi-branch switching in the inverse kinematics solution of the three-axis parallel robot are avoided. Under the synergistic effect of the three mechanisms, the smoothness of the robot's end-effector trajectory is improved by more than 90%, which fully meets the application requirements of precision teaching. Moreover, the above effects can be stably achieved when the key point extraction algorithm of the hand skeleton is replaced with OpenPose and HRNet, the angle sensing unit 200 is replaced with a MEMS inertial measurement unit, and the industrial controller execution unit 400 is replaced with a real-time controller from other brands such as Siemens and Beckhoff.

[0027] In one possible implementation, step S2 involves obtaining the first coordinate information within the robot's working plane through normalized coordinate mapping, specifically including: The host computer data processing unit 300 converts the normalized coordinates into actual pixel coordinates based on the image width and height of the vision acquisition unit 100; Then, a camera plane coordinate system was established with the image center as the origin, and the pixel coordinates of the wrist key points were converted into camera plane coordinates (x, y, z). c , y c ); Through perspective transformation matrix H m Map the camera plane coordinates to the robot's working plane coordinates, and set the homogeneous form of the transformed camera plane coordinates as follows: The homogeneous coordinates after perspective transformation are The robot's working plane coordinates satisfy the following perspective relationship: , The perspective relationship is used to eliminate perspective distortion caused by the tilted installation of the camera; Where, x r and y r Δx and Δy are the X-axis and Y-axis coordinates in the robot's working plane, respectively, and Δx and Δy are the adjustment offsets in the X-axis and Y-axis directions, respectively. and These are the first and second components of the homogeneous coordinates after perspective transformation, respectively. The scaling factor is dynamically adjusted as the camera's planar coordinate position changes.

[0028] This is a vision-angle meter fusion servo control solution specifically designed for non-contact teaching scenarios of three-axis Delta parallel robots. Its core purpose is to address the technical pain points of existing monocular vision systems, such as the inability to stably acquire depth information, the high cost of depth camera hardware, and trajectory step jitter caused by discrete gesture control. The overall system constructs a closed-loop control chain consisting of a perception layer, a processing layer, and a control layer. The perception layer includes a vision acquisition unit 100 for acquiring 60° images of the operator's hand and an angle sensing unit 200 fixed to the operator's wrist to acquire pitch angles. The vision acquisition unit 100 can be an industrial CCD camera calibrated with internal parameters. Machines, CMOS industrial cameras, global shutter network cameras, or embedded vision modules equipped with OV5640 sensors can be used to adapt to dynamic teaching requirements, as long as the acquisition frame rate is not less than 25fps and the resolution is not less than 640×480. The angle sensing unit 200 can be a DataLITE wireless angle sensor, a MEMS three-axis inertial measurement unit such as MPU6050 or BMI088, a patch tilt sensor, or even a Hall sensor with angle output. The range can cover ±90°, the accuracy is not less than 0.1°, and the weight can be controlled within 50g to avoid interfering with the operator's hand dexterity. The processing layer is a host computer data processing unit 300 equipped with a Python runtime environment. It can be equipped with devices with floating-point computing capabilities, such as industrial control computers, high-performance desktop PCs, Raspberry Pi 4B, NVIDIA Jetson Nano, or Rockchip RK3588 embedded development boards. It only requires a minimum of four cores with a 1.5GHz main frequency to support the algorithm operation. The control layer consists of an industrial controller execution unit 400, which can be a soft PLC based on the CODESYS platform, a Siemens S7-1200 / 1500 series PLC, a Beckhoff CX series embedded controller, an Omron NX series controller, or a motion control board with an STM32H743 chip. It only needs to support a PLC scan cycle of less than 1ms and have real-time bus interfaces such as EtherCAT and PROFINET. The modules are connected via adapter links. The vision acquisition unit 100 interacts with the host computer via USB 3.0, Gigabit Ethernet, or MIPI-CSI interfaces. The angle sensing unit 200 communicates with the host computer via Bluetooth 5.0, WiFi 6, or a 2.4G wireless serial port. The host computer and the industrial controller transmit data bidirectionally via industrial Ethernet protocols such as UDP, TCP, or Modbus TCP. The industrial controller communicates via EtherCAT, PROFINET, PowerLink, or SERCOS. The III-type real-time bus is electrically connected to the three servo drives of the three-axis parallel robot. All links can be replaced with wired or wireless forms according to the on-site deployment requirements. During operation, the host computer first calls any hand skeleton key point extraction algorithm such as MediaPipe Hand Model, OpenPose, or HRNet to identify the operator's hand skeleton, extract the normalized coordinates of the wrist key points, and convert the normalized coordinates into actual pixel coordinates based on the image width and height of the vision acquisition unit 100. Then, a camera plane coordinate system is established with the image center as the origin, and the pixel coordinates of the wrist key points are converted into camera plane coordinates (xc, yc). The camera plane coordinates are mapped to the robot's working plane coordinates through a 3×3 perspective transformation matrix Hm obtained in advance using the checkerboard calibration method. The homogeneous form of the converted camera plane coordinates is set as follows: The homogeneous coordinates after perspective transformation are Then the robot's working plane coordinates are (x r , y r The robot's working plane coordinates satisfy the following perspective relationship: , The perspective relationship is used to eliminate perspective distortion caused by the tilted installation of the camera, wherein and These are the X-axis and Y-axis coordinates in the robot's working plane, respectively. and The adjustment offset is adaptively adjusted based on the operator's posture and arm length. This mapping process can also be adapted to a coordinate system with the top left corner of the image as the origin, requiring only synchronous adjustment. , The value can be chosen freely, and the homogeneous form can be adjusted to column vector or row vector form according to the matrix operation requirements. The core is to compensate for the mapping error caused by camera mounting distortion and field of view offset through perspective transformation. The reasoning logic based on the above structure is as follows: In the existing technology, when mapping planar coordinates with monocular vision, simple linear proportional conversion is often used, which cannot compensate for the error caused by camera mounting tilt and perspective distortion. Moreover, coordinate system alignment requires a complex multi-parameter calibration process, which is difficult to deploy and has low accuracy. However, this solution reduces the mapping error of XY coordinates from ±5mm in the existing technology to within ±0.5mm through the correction mechanism of perspective transformation matrix. At the same time, a camera coordinate system is established with the image center as the origin and the Y-axis is negative to correct the direction. Only two adjustment offset values ​​are needed to complete the cross-coordinate system alignment, which greatly reduces the deployment difficulty. Furthermore, the perspective transformation matrix Hm can be flexibly obtained through the conventional checkerboard calibration method without fixed values. It can be adapted to various deployment scenarios such as camera tilt installation, side mounting, and inverted installation. The source of normalized coordinates can also be replaced by the output of any hand key point extraction algorithm. All features can be adjusted according to actual needs, fully verifying the replaceability of relevant features. The resulting benefits are as follows: the planar coordinate mapping method of this solution can still stably achieve a coordinate mapping accuracy within ±1mm even when the visual acquisition unit 100 is replaced with a regular high-definition network camera, the hand key point extraction algorithm is replaced with OpenPose, and the origin of the coordinate system is adjusted to the upper left corner of the image. This effectively solves the pain points of low accuracy and difficult deployment of existing monocular vision coordinate mapping, and also lays a reliable planar coordinate foundation for subsequent three-dimensional coordinate fusion and smooth follow-up. At the same time, all replaceable embodiments of refined features are clearly described in the specification, fully supporting the superior generalization of the claims, which complies with the provisions of Article 26, Paragraph 4 of the Patent Law that the claims should be based on the specification. Moreover, for the high-speed and lightweight characteristics of the three-axis parallel robot, high-precision planar coordinate mapping can effectively avoid the cumulative error of subsequent inverse kinematics solution, further improve the smoothness of the follow-up trajectory, solve the trajectory step problem caused by discrete gesture control, and finally achieve low-cost, high-precision non-contact continuous teaching effect.

[0029] In one possible implementation, step S2 converts the pitch angle into the second coordinate information of the robot end effector via a nonlinear mapping model, specifically including: The host computer data processing unit 300 calls the online interface of the angle sensing unit 200 to read the raw return value D. raw+4000 D raw This is the raw digital value returned by the angle sensing unit through the online interface; Since the original return value contains a fixed bias of 4000, bias correction is performed first: D adc =D raw+4000 -4000; Among them, D adc These are the sensor values ​​after bias correction; Then based on the proportionality coefficient k d D adc Converted to initial angle value θ d θ d =D adc ·k d ; Where, θ d k is the initial angle value output by the angle sensing unit 200. d Let k be the sensor proportionality coefficient. d =0.045; For θ d Perform safety limiting to obtain the current pitch angle θ c ; After the amplitude is limited, θ c Substituting into the quadratic mapping model, the absolute Z-axis coordinate z of the robot's end effector is calculated. t :z t =−0.0030219780·θ c 2 −0.9313186813·θ c -420.0; Among them, z t To obtain the target Z-axis coordinates of the three-axis parallel robot end effector through mapping, this quadratic mapping model is obtained by fitting three sets of calibration constraint points: when θ c z =−60° t =−375mm, when θ c =0° z t =−420mm, when θ c =70° z t =−500mm.

[0030] This step is the core sub-process for generating the absolute Z-axis coordinates of the robot's end effector in the vision-angle meter fusion servo control method. It is implemented collaboratively by the angle sensing unit 200 and the host computer data processing unit 300. The angle sensing unit 200 can be replaced by any one of the following: a DataLITE wireless angle sensor, a MEMS triaxial inertial measurement unit such as MPU6050 / BMI088, an SCA103T piezoelectric tilt sensor, or a flexible strain gauge angle sensor, as long as it meets the requirements of a range coverage of ±90° and a measurement accuracy ≥0.1. The sensor's inherent bias is 2048, and its output interface is compatible with any wired or wireless communication protocol, such as serial port, I2C, SPI, Bluetooth 5.0, and WiFi 6. Different sensors have different inherent biases for their raw return values, not just 4000. For example, the inherent bias of a 12-bit ADC output sensor is 2048, and that of a 16-bit ADC output is 32768. After the host computer calls the corresponding interface to read the raw value, it first performs bias correction to eliminate zero drift and obtains the bias-corrected sensor value. Then, based on the sensor's... The proportional coefficient obtained from calibration, within the range of 0.02 to 0.1, is converted into the initial angle value. The specific value of the proportional coefficient is adjusted according to the sensor model. For example, when using MPU6050, the proportional coefficient can be 0.038. Then, a safety limit is applied to the initial angle value. The limit range can be adjusted according to the operator's natural wrist movement range, and is not limited to [-60°, 70°]. After limiting, the current pitch angle is obtained. The current pitch angle is then input into a nonlinear mapping model. This model is not limited to the quadratic polynomial form, but can also use a cubic polynomial, piecewise linear function, or other fitting forms. Its coefficients are obtained by fitting multiple sets of calibration constraint points based on the reachable workspace of the three-axis parallel robot end effector, the safe height range of the end effector, and the comfortable range of the operator's hand movements. The number and value of the calibration constraint points can be adjusted according to the robot specifications. For example, -90° can correspond to -300mm, 0° to -450mm, and 80° to -650mm can be set as calibration points. The fitted Z-axis coordinate is the absolute height coordinate in the robot's base coordinate system, without relying on an external reference system. Finally, it is output to the three-dimensional coordinate fusion stage. In existing technologies, monocular vision calculates depth by changing the area of ​​a hand image. This is significantly affected by light intensity, hand posture, and occlusion, and the depth data can fluctuate by up to ±20mm, which cannot meet the accuracy requirements of teaching a three-axis parallel robot. While depth cameras can improve depth accuracy, their hardware cost is 3 to 5 times that of monocular solutions, their calibration process is complex, and they have stringent requirements for ambient light and the texture of the object being measured, making them unsuitable for rapid deployment in industrial settings. This step uses a low-cost wearable angle sensing unit 200 to obtain the wrist pitch angle. The physical stability of angle measurement is far superior to that of visual estimation solutions, with data jitter controlled within ±0.1°, ensuring the reliability of depth information from the source. The bias correction, scaling conversion, and safety limiting in the preprocessing stage are common operations for sensor data processing, which can be adapted to the characteristics of various angle sensing units 200, without the need for customized development for specific sensors, thus reducing the difficulty of system adaptation. Based on a nonlinear mapping model fitted with multiple calibration points, the nonlinear relationship between the wrist pitch angle and the Z-axis displacement of the robot end effector can be accurately matched. Compared with the ±2mm error of linear mapping, the fitting error of quadratic polynomial mapping can be controlled within ±0.3mm, significantly improving the accuracy of Z-axis coordinate generation. This design achieves a Z-axis coordinate repeatability error of ≤±0.5mm while reducing hardware costs by more than 80%. It is compatible with three-axis parallel robots of different specifications and can operate stably in complex industrial environments such as strong light, weak light, and dust. It effectively avoids the jitter defects of visual depth estimation, providing a reliable depth benchmark for subsequent 3D coordinate fusion and continuous servo control, solving the pain points of high cost and poor stability of existing non-contact teaching solutions for depth acquisition.

[0031] In one possible implementation, step S3 performs a first-order low-pass filtering on the fused 3D coordinates. Specifically, a first-order low-pass filtering algorithm is used, and the filtering formula for any coordinate quantity q along any coordinate axis is: q f (k)=q f (k−1)+α·[q(k)−q f [(k−1)]; Where q(k) is the original input coordinate value at the current time, q f (k) represents the filtered output coordinates at the current time, q f (k−1) represents the filtered output coordinate value at the previous time step, α is the filtering coefficient, and 0<α<1; The three-dimensional coordinates include the X-axis coordinate (x), Y-axis coordinate (y), and Z-axis coordinate (z). The above filtering operation is performed independently on x, y, and z to obtain the filtered coordinates x. f (k), y f (k), z f (k).

[0032] Specifically, in step S3, the first-order low-pass filtering process is applied to the fused three-dimensional coordinates. This is achieved using a first-order low-pass filtering algorithm. The three-dimensional coordinates include the X-axis coordinate (x), Y-axis coordinate (y), and Z-axis coordinate (z). Filtering operations are performed independently on x, y, and z, and the filtering formula is as follows: in, , , The original X-axis, Y-axis, and Z-axis coordinate values ​​at the current moment. , , These represent the filtered output X-axis, Y-axis, and Z-axis coordinates at the current time. , , The X, Y, and Z coordinates of the output after filtering at the previous time step are α, where α is the filtering coefficient and 0 < α < 1.

[0033] This step is the post-processing stage of the three-dimensional coordinates in the vision-angle sensor fusion three-axis parallel robot servo control method. It is deployed within the host computer data processing unit 300, which can be mounted on any hardware carrier with floating-point computing capabilities, such as an industrial control computer, an embedded ARM development board, or an edge computing device. It interacts with the vision acquisition unit 100, the angle sensing unit 200, and the industrial controller execution unit 400 via communication links. This stage incorporates three independent recursive first-order low-pass filter channels, corresponding to the X, Y, and Z axis components of the three-dimensional coordinates, respectively. These channels operate in parallel without signal coupling. The single-channel filtering logic is as follows: It retrieves the original input coordinate value along the corresponding coordinate axis at the current moment, reads the filtered output value of the channel from the previous moment stored in the storage medium, and combines it with the preset filtering coefficients in the register. Then, it calculates the filtered output value at the current moment according to the recursive formula. The entire operation only includes three basic arithmetic operations: addition, subtraction, and multiplication, without high-complexity operations such as convolution, integration, or matrix inversion. The original input coordinate value can be adapted to coordinate components in any coordinate system, such as rectangular, cylindrical, or spherical coordinates. The filtering coefficients can be flexibly configured within the open range of 0.05 to 0.5 based on the sampling frequency and noise distribution characteristics of the sensing unit. The storage medium for the filtered output value from the previous moment can be any storage medium with single-cycle read / write capability, such as the processor's on-chip register, double-rate random access memory, or non-volatile flash memory. The recursive form of the filtering formula can be replaced with an iterative equivalent form, as long as the operational logic remains equivalent.

[0034] In the existing technology, the planar coordinates output by the visual recognition algorithm are subject to high-frequency jitter due to image noise and hand tremors. The pitch angle output by the angle sensing unit 200 is subject to random drift due to device thermal noise and electromagnetic interference. Coupled with the instantaneous packet loss and jump of the communication link, the unconditioned three-dimensional coordinate input will trigger the high-frequency response of the three-axis parallel robot 500 servo system, resulting in end effector tremor and trajectory glitch. If advanced filtering algorithms such as moving average and Kalman filtering are used, the single-axis single-operation time generally exceeds 10μs, which cannot meet the real-time requirement of 1ms-level control cycle for parallel robots; If a multi-axis coupled filtering scheme is used, abnormal jumps in a single axis will be transmitted to the other axes through the coupling relationship, amplifying the coordinate calculation error. This step adopts a recursive first-order low-pass filter design with independent channels. The single-axis single-operation time is less than 1μs, which can adapt to the real-time processing requirements of various computing platforms. The independent filtering design can adjust the filter coefficients separately for the noise characteristics of different coordinate components, balancing response speed and filtering effect, and avoiding the transmission of errors across axes.

[0035] This design can suppress the jitter amplitude of three-dimensional coordinates to within ±0.2mm. Compared with the unfiltered solution, the smoothness of the robot's end-effector trajectory is significantly improved. Compared with the second-order filtering solution, the computation time is lower, effectively avoiding servo oscillation caused by high-frequency noise, while avoiding follow-up lag caused by filtering delay. It is suitable for the high-speed motion characteristics of three-axis parallel robots. Moreover, the filtering algorithm, coordinate component type, filtering coefficient, and storage medium can be flexibly replaced according to the deployment scenario, and it is compatible with different specifications of sensing units and control platforms.

[0036] In one possible implementation, the workspace validity determination in step S3 specifically includes: Based on the filtered plane coordinates (x) f (k), y f (k) Calculate the plane radius r of the target point relative to the center of the workspace: ; Set the maximum allowable radius r of the workspace max ; When r≤r max And the Z-axis coordinate z f (k) When within the preset safety height range, generate the validity flag bit valid=1; Otherwise, the validity flag is set to valid=0, prohibiting the robot from moving.

[0037] This step is a preliminary safety protection step in the non-contact servo control process of a three-axis parallel robot. It is integrated into the monitoring logic of the host computer data processing unit 300 or the industrial controller execution unit 400 and is used to perform real-time compliance verification of the filtered 3D target coordinates. This step first calculates the Euclidean distance from the target point to the center of the workspace based on the filtered plane coordinates in the Cartesian coordinate system through the arithmetic operation unit to obtain the plane radius. The operation unit can be adapted to any hardware carrier among the central processing unit, digital signal processor, field-programmable gate array, or application-specific integrated circuit. The source of the plane coordinates can also be replaced by the radial component in polar coordinates or the projection value obtained by cylindrical coordinate transformation. Subsequently, the logic judgment unit compares the calculated planar radius with the preset maximum allowable radius using a threshold, and simultaneously performs an inclusion judgment between the filtered Z-axis coordinate and the preset safe height range. The maximum allowable radius value can be adaptively set according to the link length parameters and workspace envelope of the three-axis parallel robot, and the safe height range can also be dynamically adjusted according to the end effector type and worktable height. In existing technologies, non-contact teaching systems often cause the target point to exceed the robot's physical travel distance due to operator error or visual recognition drift, easily triggering mechanical hard limit collisions or motor stall alarms. Moreover, simple boundary protection is insufficient to address singularity issues in the Z-axis direction. This step constructs a dual constraint mechanism of planar radius and Z-axis height, using simple geometric calculations to determine in real time whether the target point is within the reachable workspace. Compared to complex six-dimensional spatial envelope algorithms, the computational latency is reduced to the microsecond level, meeting real-time control requirements. The binary validity flag generated by the logic judgment unit can be directly connected to the enable link of the industrial controller. When valid=0, the servo pulse output is forcibly cut off or the brake is triggered, preventing over-limit movement from the decision-making source. This design effectively avoids the risk of robot overtravel collisions without adding additional hardware sensors, improving the safety of human-robot collaborative operations. Moreover, the radius calculation method, threshold setting logic, and flag control strategy can all be equivalently replaced according to the specific robot configuration and control architecture, demonstrating strong adaptability.

[0038] In one possible implementation, step S5 involves latching the received three-dimensional target position according to a preset latching period, specifically including: Set the latch frequency division factor N and the PLC scan cycle T of the industrial controller. s Determine the target point latching period T l :T l =N·T s ; Among them, T s The PLC scan cycle is N, where N is the latch frequency division coefficient, and T is T. l The target point latching period; Only in each latching cycle T l Upon arrival, the host computer data processing unit 300 will send the latest three-dimensional target position (x) net (k), y net (k), z net (k) is updated to the target coordinates (x) used in the current inverse kinematics calculation. cmd (k), y cmd (k), z cmd (k)), keeping the target coordinates constant during the latching cycle interval; Where, x net (k), ynet (k), z net (k) represent the target positions of the X, Y, and Z axes sent by the host computer data processing unit during the k-th latching cycle, respectively. cmd (k), y cmd (k), z cmd (k) represents the target coordinates of the X-axis, Y-axis, and Z-axis used for inverse kinematics calculation after being latched by the industrial controller execution unit in the k-th latching cycle.

[0039] This step is the target coordinate timing control link in the non-contact follow-up control process of the three-axis parallel robot. It belongs to the real-time task scheduling architecture of the industrial controller execution unit 400 and includes a three-layer structure from top to bottom: periodic parameter configuration unit, latch trigger unit, and double buffer coordinate storage unit. The cycle parameter configuration unit can be integrated into the software PLC configuration interface, hardware PLC system registers, and embedded controller non-volatile memory area, supporting online adjustment of the latch frequency division coefficient N and PLC scan cycle T. s The PLC scan cycle T s It can be configured to values ​​such as 1ms, 2ms, 4ms, and 8ms to adapt to the controller's computing capabilities. The latch division coefficient N can be configured to an integer between 10 and 100 to match the visual data update frequency. The latch trigger unit can be adapted to the controller's cyclic interrupt organization block, hardware timer overflow signal, external IO rising edge trigger logic, or can be replaced by DMA transfer completion interrupt and other event triggering mechanisms. The double-buffered coordinate storage unit includes a network receive buffer and a reverse decoding instruction buffer. The former is used to temporarily store the latest three-dimensional target position (x, y, y) sent by the host computer data processing unit 300. net (k), y net (k), z net (k)), the latter is used to store the target coordinates (x) called by the current inverse kinematics solving unit. cmd (k), y cmd (k), z cmd (k)) The storage medium can be on-chip RAM, dual-port SRAM, FPGA distributed registers and other carriers with single-cycle read and write capabilities.

[0040] In the existing technology, the typical output frame rate of the vision acquisition unit 100 is 25~30fps, corresponding to a data update cycle of 33~40ms. However, the data transmission cycle of the host computer data processing unit 300 and the PLC scanning cycle of the industrial controller are usually 1~4ms. There is a significant timing asynchronous problem between the two. Without a timing control mechanism, the controller will read data from the network receiving buffer in each scanning cycle. When the visual data has not been updated, it is easy to read residual old data or abnormal values ​​caused by communication noise, which will cause the target coordinates of the inverse kinematics input to change frequently. The three-axis parallel robot 500, due to its high-speed and lightweight dynamic characteristics, is far more sensitive to target coordinate changes than a serial robot. Even minor coordinate changes can trigger high-frequency end effector vibration, trajectory glitches, or even servo overload, failing to meet the smoothness requirements of the taught trajectory. This step uses the controller's inherent PLC scan cycle T. s Using the reference clock as a reference, the reference clock is divided by the latch division factor N to obtain the target point latching period T that matches the visual data update frequency. l =N·T s Only in T l The trigger edge performs a coordinate copy operation, transferring the latest three-dimensional coordinates of the network receive buffer to the inverse decoding instruction buffer. The coordinates of the inverse decoding instruction buffer remain constant during the latch interval, thus blocking the coordinate jump input caused by high-frequency scanning at the timing level.

[0041] This design eliminates the need for additional hardware timers or external synchronization modules, directly reusing the controller's native scan cycle resources, resulting in extremely low hardware costs. The latching period can be flexibly adapted to the output frame rate of different vision devices by adjusting the frequency division coefficient N, taking into account both the follow-up response speed and trajectory smoothness. The dual-buffered storage mechanism avoids coordinate read / write conflicts and is compatible with various control platforms such as soft PLCs, hard PLCs, and embedded motion controllers. After adopting this latching mechanism, the trajectory jitter amplitude of the end effector of the three-axis parallel robot is reduced, the servo current fluctuation amplitude is reduced, and the control instability problem caused by asynchronous updates of visual data is effectively avoided, adapting to the human-machine collaborative safety requirements in non-contact teaching scenarios.

[0042] In one possible implementation, the PLC scan cycle T s The latch time is 4ms, the latch frequency divider N is 25, and the target point latch period T is... l It takes 100ms.

[0043] This preferred embodiment is a concrete implementation of the aforementioned target coordinate latching mechanism. It belongs to the real-time task scheduling architecture of the industrial controller execution unit 400. The architecture is layered from top to bottom: the top-level industrial controller can be replaced by any control device with real-time scheduling capabilities, such as CODESYS soft PLC, Siemens S7-1500 / 1200 series hard PLC, Beckhoff CX2040 embedded controller, self-developed motion control board based on STM32H743 chip, or FPGA architecture dedicated motion controller. The middle layer is the parameter configuration unit, and the periodic reference source is the system timer interrupt of the controller, with a scan period T. s It can be configured to steady-state periods natively supported by the controller, such as 1ms, 2ms, 4ms, and 8ms. The latching frequency division coefficient N is stored in the controller's non-volatile register, the host computer configuration interface, or the Flash storage area, and can be any integer between 10 and 100. The underlying layer consists of latch triggering and data interaction units. The latch triggering logic can be adapted to the controller's cyclic interrupt organization block and hardware timer overflow signals. The dual-buffered storage area can utilize on-chip RAM, dual-port SRAM, FPGA distributed registers, or other storage media with single-cycle read / write capabilities. In existing technologies, the output frame rate of the vision acquisition unit 100 is mostly 25fps (corresponding to an update cycle of approximately 40ms), which is significantly asynchronous with the 1-4ms scanning cycle of the industrial controller. If the latch interval is too short, it is easy to read dirty intermediate data that has not been filtered by the host computer, causing coordinate jumps; if it is too long, it will cause lag in follow-up and make it impossible to reproduce hand movements. This solution selects 4ms as the PLC scanning cycle, which is the native steady-state parameter of mainstream industrial controllers, and can operate stably without modifying the underlying firmware. A latching frequency division coefficient of 25 was selected, and an integer division yielded a target point latching period of 100ms. This not only covers the 2-3 frame data window of the visual output to filter out abnormal jumps in a single frame, but also matches the response speed requirements of non-contact teaching to avoid lag. This parameter combination does not require the addition of a hardware synchronization module, directly reusing the controller's native scheduling resources. The integer frequency division logic does not require complex decimal calculations, the controller's CPU utilization rate is not significantly changed, and it does not crowd out the computing power for inverse kinematics solving and servo closed-loop control. Under these parameters, the packet loss rate of visual data transmission is less than 0.1%, the frequency of coordinate jumps is reduced to 0 times / second, and the end-effector following error of the three-axis parallel robot is stably controlled within 1mm. Furthermore, the parameters are universally applicable, including T. s Adjusting Ts to 2ms and N to 50, or Ts to 8ms and N to 12, will yield similar latching periods, adapting to control devices of different brands and architectures.

[0044] In one possible implementation, step S5 converts the target joint angle into a servo position command, specifically including: Set the mounting direction coefficient d for each active branch servo axis i With zero-point offset o i , where i=1,2,3 correspond to the three active branches of the three-axis parallel robot; The position of each active arm in the active chain when it is in a horizontal state is set as the mechanical zero position of the corresponding active joint, at which time the corresponding active joint angle is 0°. During the k-th latching cycle, the target joint angle θ of the i-th active branch is...i (k) is the directed angle between the axis of the active arm of the active chain and the horizontal reference line in its rotation plane; When the active arm is in a horizontal position, θ i (k) = 0°, when the active arm rotates downward from the horizontal position, θ i (k) is a positive value, and when the active arm rotates upward from the horizontal position, θ i (k) is a negative value; According to the formula, the target joint angle θ of the i-th axis is... i (k) is converted to the servo position command p of the i-th axis. i (k): p i (k)=d i ·θ i (k)+o i ; Where k is the latch cycle number, p i (k) represents the target position command of the i-th servo axis in the k-th latching cycle, d i Let be the servo direction coefficient for the i-th axis, used to correct the consistency between the forward rotation direction of the servo motor and the motion direction of the robot's end effector. i This is used to offset the zero point of the i-th axis, and is used to compensate for the deviation between the servo axis zero point and the robot's mechanical zero position. And it is sent to the corresponding servo driver via the real-time bus.

[0045] This step is the instruction adaptation stage in the non-contact follow-up control process of a three-axis parallel robot, where the inverse kinematics output is transmitted to the servo drive. It is deployed in the servo instruction generation unit of the industrial controller execution unit 400. The architecture is layered from top to bottom: In the top-level control link, the industrial controller can be replaced by a CODESYS soft PLC, a Siemens S7-1200 / 1500 series hard PLC, a Beckhoff CX2040 embedded controller, a self-developed motion control board based on an STM32H743 or DSP28335 chip, a dedicated motion controller based on an FPGA architecture, or other control devices with real-time floating-point computing capabilities. The servo drive can be adapted to any three-axis Delta matching servo product that supports position modes, such as the Panasonic MSMF series, Mitsubishi HG-KR series, or Inovance IS620P series. The intermediate instruction conversion unit embeds a parameter storage submodule and a computation submodule. The storage medium of the parameter storage submodule can be replaced by the controller's on-chip EEPROM, ferroelectric memory, non-volatile Flash, or host computer configuration database, used to store the installation direction coefficients d of each active branch. i With zero-point offset o i The computational submodule can be adapted to CPU arithmetic logic units, DSP multiply-accumulate units, and FPGA hard-core multipliers to perform linear coordinate transformation operations. In the underlying data transmission link, the real-time bus can be replaced with industrial bus protocols such as EtherCAT, PROFINET, PowerLink, SERCOSIII, CANopen, and Modbus TCP, which meet transmission delays of less than 1ms. It can also be adapted to the PWM / pulse sequence transmission logic of pulse-type servos. In specific implementation, indices i=1, 2, and 3 correspond to the first, second, and third active branches of the three-axis parallel robot, respectively, and d... i The value range is +1 or -1, which is used to match the consistency between the forward rotation direction of the servo motor and the motion direction of the robot end effector. For example, the forward rotation of the first axis of a Delta robot corresponds to the contraction of the active branch, while the inverse kinematics model requires that the forward rotation corresponds to the extension of the branch. Configuring d1=−1 can complete the direction correction. o i The deviation compensation value between the physical zero point of the servo axis and the theoretical mechanical zero point of the robot can be obtained through manual teaching to zero or laser interferometer calibration. The unit can be angle, radian or servo encoder unit. For example, if the mechanical zero point of the first axis servo is offset by 12.3° from the theoretical zero point, configuring o1 with the corresponding value can eliminate the assembly deviation.

[0046] In the existing technology, the servo mounting direction of a three-axis parallel robot varies depending on the overall structural design. Directly issuing the theoretical joint angles from the inverse kinematics output can easily lead to the problem that the servo steering direction is opposite to the end effector motion direction, causing the risk of collision. The physical zero point of the servo drive is affected by the assembly error of the linkage and the installation deviation of the lead screw, and there is an inherent offset from the mechanical zero point of the robot's theoretical model. Directly issuing the theoretical joint angle will cause the end-positioning accuracy to deteriorate to more than ±2mm. Some solutions embed the direction correction and zero-point compensation logic into the controller firmware, requiring code recompilation when changing robot models or servo configurations, resulting in high adaptation costs. This solution uses linear transformation logic containing only multiplication and addition, with a single-axis operation time of less than 1μs, without consuming computing resources in the control cycle, and is suitable for high-speed three-axis Delta millisecond-level control requirements; d i With o i With configurable parameters, it can be adapted to Delta robots of different configurations and servo drives of different brands without modifying the underlying firmware. It can even be extended to servo command conversion for Scara robots and serial robots, significantly improving its versatility. Direction coefficient correction completely eliminates the risk of servo steering errors, and zero-point offset compensation can improve end-positioning accuracy to within ±0.5mm; The parameters are stored on non-volatile media, are not lost when power is lost, and do not need to be recalibrated after power is restored, thus improving deployment efficiency; The conversion logic can be equivalently replaced by the instruction generation form in pulse mode, which is compatible with pulse-type servo drive solutions and is not limited to bus-type control architecture.

[0047] In one possible implementation, when the industrial controller execution unit 400 calls the inverse kinematics solving unit, it uses the legal joint angle calculated in the previous latching cycle as the initial reference value for the current inverse kinematics operation to reduce motion discontinuity caused by inverse kinematics branch switching.

[0048] This step is the initialization configuration stage for inverse kinematics solution in the non-contact servo control process of a three-axis parallel robot. It is integrated into the inverse kinematics solution unit of the industrial controller execution unit 400. The architecture is layered from top to bottom: the top-level industrial controller can be replaced by CODESYS soft PLC, Siemens S7-1200 / 1500 series hard PLC, Beckhoff CX2040 embedded controller, self-developed motion control board based on STM32H743 or DSP28335 chip, FPGA architecture dedicated motion controller and other control devices with real-time computing capabilities. The mid-level inverse kinematics solving unit can be adapted to any computational carrier that can output the target joint angle of the three-axis Delta active branch, such as analytical inverse kinematics firmware, Newton-Raphson iterative solution module, joint space lookup table interpolation module, and lightweight neural network inverse kinematics model. The underlying initial reference value cache unit serves as the pre-storage module for inverse kinematics operations. The storage medium can be replaced with on-chip static random access memory, dual-port synchronous static random access memory, or ferroelectric non-volatile memory. It stores the valid joint angles from the previous latching cycle, verified by workspace validity and inverse kinematics convergence. These valid joint angles are those that have passed joint soft and hard limit constraints and have no risk of exceeding limits. Indices i=1, 2, and 3 correspond to the three active branches of the three-axis parallel robot. The inverse kinematics of a three-axis parallel robot exhibits multi-branch solution characteristics. Existing inverse kinematics solutions often use fixed theoretical zero points, random initial values, or current joint feedback values ​​as the initial inverse kinematic reference. When continuously input target coordinates cross the convergence domains of different inverse kinematic branches, it easily triggers inverse kinematic branch switching, causing a step jump in the target joint angle of the active branch, leading to discontinuous robot end-effector motion, trajectory spikes, or even servo overload. If an iterative inverse kinematics algorithm is used, improper selection of initial values ​​can lead to slow convergence and excessive computation time, making it unsuitable for millisecond-level control cycle requirements. This solution uses the constant target coordinates within the latching cycle as the computation benchmark and takes the valid joint angle verified by the entire link in the previous latching cycle as the initial reference value for the current inverse kinematics operation. Since the operator's hand movements are continuous in the servo scenario, the change in target coordinates between adjacent latching cycles is minimal. The valid joint angle of the previous cycle naturally falls within the convergence domain of the inverse kinematics branch corresponding to the current target coordinates, thus avoiding branch switching caused by the initial value crossing the convergence domain. Meanwhile, the validity of joint angles has been verified in the preceding workspace validity judgment and joint amplitude limiting steps, so there is no need to add extra validity verification overhead and no risk of inverse solution divergence caused by illegal initial values ​​will be introduced. This logic only requires reading the cached valid joint angles as initial input before the inverse kinematics operation, without adding any new hardware modules or complex calculations, and is compatible with the operating characteristics of various inverse kinematics algorithms. This design can eliminate joint angle jumps caused by inverse kinematics branch switching, improve the smoothness of the end effector motion, and enhance trajectory continuity; For iterative inverse kinematic algorithms, the computation time required for iterative convergence can be shortened, the computing power of the controller can be reduced, and it can be adapted to a 1ms-level high-speed control cycle. The cached valid joint angles can be stored on non-volatile media before power failure, and there is no need to re-initialize the reverse engineering parameters after power is restored, thus improving deployment efficiency. This logic can be equivalently adapted to other parallel robot configurations with inverse multi-branch characteristics, such as the Scara robot and the Stewart six-DOF parallel platform, demonstrating strong versatility.

[0049] In one possible implementation, establishing a communication connection in step S4 specifically includes: the industrial controller execution unit 400 sending a vision enable signal to the host computer data processing unit 300 through the communication channel; The host computer data processing unit 300 starts or stops the image acquisition and coordinate calculation process according to the vision enable signal; The host computer data processing unit 300 encapsulates and sends data frames in a fixed data format, which is "X,Y,Z,valid". Where X is the X-axis coordinate of the robot's working plane, Y is the Y-axis coordinate of the robot's working plane, Z is the absolute Z-axis coordinate of the robot's end effector, and valid is the validity flag.

[0050] This step is the cross-module interactive control link of the vision-angle meter integrated three-axis parallel robot non-contact follow-up control system. It is used to realize bidirectional collaboration between the industrial controller execution unit 400 and the host computer data processing unit 300. The architecture is layered from top to bottom: the top interactive main body includes the industrial controller execution unit 400 and the host computer data processing unit 300. The former can be replaced by any control device with real-time communication and scheduling capabilities, such as CODESYS soft PLC, Siemens S7-1200 / 1500 series hard PLC, Beckhoff CX2040 embedded controller, self-developed motion control board based on STM32H743 or DSP28335 chip, FPGA architecture dedicated motion controller, etc. The latter can be replaced by industrial control computer with x86 / ARM architecture processor, NVIDIA Jetson Nano edge computing device, Raspberry Pi 4B embedded development board, FPGA-based image processing acceleration card, etc., which have visual computing capabilities. The mid-level communication channel can be adapted to industrial Ethernet protocols such as UDP, TCP, and Modbus TCP, or can be replaced with fieldbuses such as RS485 and CANopen, or wireless transmission links such as WiFi 6 and Bluetooth 5.0, as long as it meets the performance requirements of end-to-end transmission latency of less than 10ms and packet loss rate of less than 0.5%. The underlying interaction logic is divided into two sub-modules: downlink power control and uplink data encapsulation. The downlink power signal can be replaced by control register bit flipping, specific string instructions, or TTL level transitions. The host computer response logic can be adapted to different control granularities, such as image acquisition process start / stop, MediaPipe / OpenPose key point extraction algorithm on / off, and dynamic adjustment of coordinate calculation frequency. The encoding format of the uplink data encapsulation can be replaced by any format agreed upon by both the sender and receiver, such as comma-separated strings, binary simplified frames, JSON structured messages, or Protobuf serialized data. The coordinate units of X, Y, and Z can be adapted to millimeters, pulse equivalents, and servo encoder counting units. The validity flag can be defined as 0 for invalid and 1 for valid, or extended to 2 for warning and 3 for fault. The data frame transmission mode can be adapted to a periodic active reporting or a query mode of controller request-host computer response.

[0051] Existing non-contact teaching systems mostly adopt an interaction mode where the host computer actively pushes data. The controller cannot actively manage the start and stop of the vision process, which easily leads to the host computer wasting computing power during teaching idle periods. Furthermore, the separate transmission of coordinate data and validity indicators increases the number of communication interactions and parsing latency. Invalid coordinates can easily trigger robot actions, posing safety hazards. Additionally, proprietary communication formats require protocol reconstruction for cross-platform integration, resulting in poor compatibility. This solution adopts a controller-led master-slave interaction logic. The industrial controller issues vision enable signals on demand, only initiating the host computer's image acquisition and coordinate calculation process during the teaching phase, reducing unnecessary power consumption. The three-dimensional coordinates and validity flags are encapsulated into a single frame of data in a fixed format, reducing the number of interactions and parsing complexity, and adapting to the real-time requirements of millisecond-level control cycles; The simplified field order conventions eliminate the need for complex encoding and decoding logic. Different software and hardware platforms can complete the integration simply by following the format conventions, reducing cross-platform adaptation costs. The validity flag is strongly bound to the coordinate data, allowing the controller to quickly determine data availability without additional verification, thus avoiding malfunctions caused by invalid coordinates at the source. This design reduces the unnecessary computing power consumption of the host computer during idle periods, reduces communication transmission latency and parsing jitter, and, in conjunction with the aforementioned latch cycle control mechanism, improves the smoothness of robot end effector motion while enhancing the safety of human-robot collaborative operations. The standardized interaction format can be reused in visual teaching scenarios for multi-configuration industrial robots.

[0052] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for remote servo control of a parallel robot based on vision-angle meter fusion, characterized in that, Includes the following steps: S1. Construct a closed-loop control system: Construct a closed-loop control system consisting of a vision acquisition unit, an angle sensing unit, a host computer data processing unit, and an industrial controller execution unit. The visual acquisition unit is configured to acquire images of the operator's hand. The angle sensing unit is configured to be fixed to the operator's wrist to collect the pitch angle; The host computer data processing unit is communicatively connected to the vision acquisition unit and the angle sensing unit. The industrial controller execution unit and the host computer data processing unit interact bidirectionally via the industrial Ethernet protocol, and its output is electrically connected to the three servo drives of the three-axis parallel robot via a real-time bus. S2. Generate three-dimensional target coordinates: The operator's hand image is acquired through the vision acquisition unit, the hand skeleton is identified by the hand skeleton key point extraction algorithm and the normalized coordinates of the wrist key points are extracted, and the normalized coordinates are mapped to obtain the first coordinate information in the robot's working plane. At the same time, the pitch angle of the operator's wrist is acquired through the angle sensing unit, and the pitch angle is converted into the second coordinate information of the robot end through a nonlinear mapping model. S3. Data frame generation: The first coordinate information and the second coordinate information are fused together in time synchronization, and the fused three-dimensional coordinates are subjected to first-order low-pass filtering and workspace validity judgment to generate a data frame containing the three-dimensional target position and validity flag. The three-dimensional coordinates include X-axis coordinate x, Y-axis coordinate y and Z-axis coordinate z. S4. Data frame transmission: Establish a communication connection between the host computer data processing unit and the industrial controller execution unit, and send the data frame to the industrial controller execution unit; S5. Latching and Inverse Kinematics Solving: In the industrial controller execution unit, the received three-dimensional target position is latched according to a preset latching period. The latched target position is input into a preset inverse kinematics solving unit to calculate the target joint angle of each active branch. The target joint angle is converted into a servo position command and sent to the servo driver to drive the robot end effector to follow the operator's hand movements.

2. The vision-angle meter fusion parallel robot remote follow-up control method according to claim 1, characterized in that, Step S2, which involves mapping the normalized coordinates to obtain the first coordinate information within the robot's working plane, specifically includes: The host computer data processing unit converts the normalized coordinates into actual pixel coordinates based on the image width and height of the visual acquisition unit; Subsequently, a camera plane coordinate system is established with the image center as the origin, and the pixel coordinates of the wrist key points are converted into camera plane coordinates (x c , y c ), wherein x c and y c are the coordinates of the X-axis and the Y-axis in the camera plane coordinate system with the image center as the origin, respectively. Through perspective transformation matrix H m The camera plane coordinates are mapped to the robot's working plane coordinates, and the homogeneous form of the transformed camera plane coordinates is set as follows: , where p c This is the homogeneous form of the camera plane coordinates with the image center as the origin; The homogeneous coordinates after perspective transformation are Then the robot's working plane coordinates are (x r , y r The robot's working plane coordinates satisfy the following perspective relationship: , The perspective relationship is used to eliminate perspective distortion caused by the tilted installation of the camera; Where, x r and y r Δx and Δy are the X-axis and Y-axis coordinates in the robot's working plane, respectively, and Δx and Δy are the adjustment offsets in the X-axis and Y-axis directions, respectively. and These are the first and second components of the homogeneous coordinates after perspective transformation, respectively. The scaling factor is dynamically adjusted as the camera's planar coordinate position changes.

3. The vision-angle meter fusion parallel robot remote follow-up control method according to claim 1, characterized in that, Step S2, which involves converting the pitch angle into the second coordinate information of the robot's end effector via a nonlinear mapping model, specifically includes: The host computer data processing unit calls the online interface of the angle sensing unit to read the raw return value D. raw+4000 D raw This is the raw digital value returned by the angle sensing unit through the online interface; Since the original return value contains a fixed bias of 4000, bias correction is performed first: D adc =D raw+4000 -4000; Among them, D adc These are the sensor values ​​after bias correction; Then based on the proportionality coefficient k d D adc Converted to initial angle value θ d θ d =D adc ·k d ; Where, θ d k is the initial angle value output by the angle sensing unit. d Let k be the sensor proportionality coefficient. d =0.045; For θ d Perform safety limiting to obtain the current pitch angle θ c ; After the amplitude is limited, θ c Substituting into the quadratic mapping model, the absolute Z-axis coordinate z of the robot's end effector is calculated. t :z t =−0.0030219780·θ c 2 −0.9313186813·θ c -420.0; Among them, z t To obtain the target Z-axis coordinates of the three-axis parallel robot end effector through mapping, this quadratic mapping model is obtained by fitting three sets of calibration constraint points: when θ c z =−60° t =−375mm, when θ c =0° z t =−420mm, when θ c =70° z t =−500mm.

4. The vision-angle meter fusion parallel robot airborne follow-up control method according to claim 3, characterized in that, Step S3 describes performing a first-order low-pass filter on the fused 3D coordinates. Specifically, a first-order low-pass filter algorithm is used. The 3D coordinates include the X-axis coordinate (x), Y-axis coordinate (y), and Z-axis coordinate (z). Filtering operations are performed independently on x, y, and z, and the filtering formula is as follows: in, , , The original X-axis, Y-axis, and Z-axis coordinate values ​​at the current moment. , , These represent the filtered output X-axis, Y-axis, and Z-axis coordinates at the current time. , , The X, Y, and Z coordinates of the output after filtering at the previous time step are α, where α is the filtering coefficient and 0 < α < 1.

5. The vision-angle meter fusion parallel robot remote follow-up control method according to claim 4, characterized in that, The workspace validity determination in step S3 specifically includes: Based on the filtered plane coordinates (x) f (k), y f (k) Calculate the plane radius r of the target point relative to the center of the workspace: ； Set the maximum allowable radius r of the workspace max ; When r≤r max And the Z-axis coordinate z f (k) When within the preset safety height range, generate the validity flag bit valid=1; Otherwise, the validity flag is set to valid=0, prohibiting the robot from moving.

6. The vision-angle meter fusion parallel robot airborne follow-up control method according to claim 2, characterized in that, Step S5, which involves latching the received 3D target position according to a preset latching period, specifically includes: Set the latch frequency division factor N and the PLC scan cycle T of the industrial controller. s Determine the target point latching period T l :T l =N·T s ; Among them, T s The PLC scan cycle is N, where N is the latch frequency division coefficient, and T is T. l The target point latching period; Only in each latching cycle T l Upon arrival, the latest three-dimensional target position (x) will be sent by the host computer data processing unit. net (k), y net (k), z net (k) is updated to the target coordinates (x) used in the current inverse kinematics calculation. cmd (k), y cmd (k), z cmd (k)), keeping the target coordinates constant during the latching cycle interval; Where, x net (k), y net (k), z net (k) represent the target positions of the X, Y, and Z axes sent by the host computer data processing unit during the k-th latching cycle, respectively. cmd (k), y cmd (k), z cmd (k) represents the target coordinates of the X-axis, Y-axis, and Z-axis used for inverse kinematics calculation after being latched by the industrial controller execution unit in the k-th latching cycle.

7. The vision-angle meter fusion parallel robot remote follow-up control method according to claim 6, characterized in that, The PLC scanning cycle T s The latch frequency division factor N is 25, and the target point latch period T is 4ms. l It takes 100ms.

8. The vision-angle meter fusion parallel robot remote follow-up control method according to claim 1, characterized in that, Step S5, which involves converting the target joint angle into a servo position command, specifically includes: Set the mounting direction coefficient d for each active branch servo axis i With zero-point offset o i , where i=1,2,3 correspond to the three active branches of the three-axis parallel robot; The position of each active arm in the active chain when it is in a horizontal state is set as the mechanical zero position of the corresponding active joint, at which time the corresponding active joint angle is 0°. During the k-th latching cycle, the target joint angle θ of the i-th active branch is... i (k) is the directed angle between the axis of the active arm of the active chain and the horizontal reference line in its rotation plane; When the active arm is in a horizontal position, θ i (k) = 0°, when the active arm rotates downward from the horizontal position, θ i (k) is a positive value, and when the active arm rotates upward from the horizontal position, θ i (k) is negative; According to the formula, the target joint angle θ of the i-th axis is... i (k) is converted to the servo position command p of the i-th axis. i (k): p i (k)=d i ·θ i (k)+o i ; Where k is the latch cycle number, p i (k) represents the target position command of the i-th servo axis in the k-th latching cycle, d i Let be the servo direction coefficient for the i-th axis, used to correct the consistency between the forward rotation direction of the servo motor and the motion direction of the robot's end effector. i This is used to offset the zero point of the i-th axis, and is used to compensate for the deviation between the servo axis zero point and the robot's mechanical zero position. And then send the data to the corresponding servo driver via the real-time bus.

9. The vision-angle meter fusion parallel robot remote follow-up control method according to claim 8, characterized in that, When the industrial controller execution unit calls the inverse kinematics solving unit, it uses the legal joint angle calculated in the previous latching cycle as the initial reference value for the current inverse kinematics operation to reduce motion discontinuity caused by inverse kinematics branch switching.

10. The vision-angle meter fusion parallel robot remote follow-up control method according to claim 1, characterized in that, Establishing a communication connection in step S4 specifically includes: the industrial controller execution unit sending a vision enable signal to the host computer data processing unit through the communication channel; The host computer data processing unit starts or stops the image acquisition and coordinate calculation process according to the vision enable signal; The host computer data processing unit encapsulates and sends the data frame according to a fixed data format, wherein the fixed data format is "X,Y,Z,valid"; Where X is the X-axis coordinate of the robot's working plane, Y is the Y-axis coordinate of the robot's working plane, Z is the absolute Z-axis coordinate of the robot's end effector, and valid is the validity flag.

Images