Multi-modal adaptive predictive control system for wafer transfer robot

Abstract

The embodiment of the specification provides a multi-modal adaptive predictive control system of a wafer conveying mechanical arm, high precision and high reliability of the whole process of wafer conveying are realized through integration of multi-modal perception, digital twin simulation and adaptive control. In the system initialization stage, the digital model is established through automatic calibration to provide an accurate benchmark for subsequent operation; the digital twin simulation predicts vibration and interference in advance, generates an optimized trajectory and a feedforward control amount, and effectively avoids operation risks; real-time adaptive motion fuses multi-source feedback and feedforward compensation to ensure trajectory tracking accuracy and actively suppress vibration; end alignment and compliant control avoid wafer damage and particle generation; health monitoring and particle control optimize motion parameters in real time to reduce pollution sources; a learning optimization mechanism continuously improves the system through data comparison and model correction, and finally improves production efficiency and product yield.

Description

Technical Field

[0001] The embodiments in this specification relate to the field of semiconductor manufacturing technology, and in particular to a multimodal adaptive predictive control system for a wafer transfer robotic arm. Background Technology

[0002] In the semiconductor manufacturing industry, traditional control systems for wafer transfer robotic arms typically rely on preset motion trajectories and fixed control parameters, lacking the ability to dynamically adapt to changes in equipment status. Existing systems often employ manual or simple automatic calibration during the initialization phase, resulting in insufficient accuracy of the kinematic model and impacting positioning accuracy. Task planning is largely based on static path calculations, failing to adequately consider the robotic arm's dynamic characteristics, such as vibration modes and interference risks, easily leading to motion instability or collisions. Real-time control often relies primarily on feedback control, making it difficult to actively suppress structural vibrations caused by acceleration and deceleration, and end effector alignment depends on coarse sensing, easily causing hard contact and particulate contamination during wafer handling. Furthermore, the system lacks continuous operational health monitoring and a self-optimization mechanism based on actual data; long-term performance degrades with equipment wear, failing to meet the ever-increasing demands for precision and reliability in high-end manufacturing.

[0003] Therefore, a better solution is urgently needed. Summary of the Invention

[0004] In view of this, embodiments of this specification provide a multimodal adaptive predictive control system for a wafer transfer robotic arm to address the technical deficiencies existing in the prior art.

[0005] According to a first aspect of the embodiments of this specification, a multimodal adaptive predictive control system for a wafer transfer robotic arm is provided, including a central collaborative decision-maker, a state perception module, a motion control module, a digital twin and predictive planning module, and an end effector module.

[0006] After the system is powered on, the central collaborative decision-maker performs initialization, drives the robotic arm to move to multiple reference calibration points distributed in the equipment chamber, collects actual position data through the state perception module and calibrates kinematic parameters and zero-position deviation, and establishes a digital model of the robotic arm. After receiving the task transmitted by the upper management system, the central collaborative decision-maker performs full-path dynamic simulation through the digital twin and predictive planning module to predict vibration modes and interference risks, and generate the optimal trajectory and feedforward control quantity. The motion control module performs adaptive motion control based on the optimal trajectory and feedforward control input, combined with the real-time status feedback from the state perception module. The end effector module uses sensors to scan the wafer pose for alignment and placement. The state perception module continuously monitors the vibration spectrum and particulate matter concentration, while the central collaborative decision-maker records the operational data and compares it with the predicted data, correcting the model parameters through learning and optimization algorithms.

[0007] In one possible implementation, during the initialization phase, the state perception module includes a vision sensor and a laser tracker to sample and measure the actual position of the robotic arm's end effector and compare it with the theoretical model to automatically calibrate the robotic arm's kinematic parameters and zero-position deviation. Simultaneously, a vibration sensor collects background noise and establishes benchmark reference data for vibration monitoring based on the collected background noise, while a force sensor performs zero-drift calibration.

[0008] In one possible implementation, during the task planning and digital twin simulation phase, the digital twin and predictive planning module calls the calibrated digital model of the robotic arm to perform full-path dynamic simulation in virtual space, calculates the vibration modes, motion time and interference risk of the robotic arm at each velocity point, generates a smooth and collision-free optimal trajectory, and calculates the feedforward control quantity required to suppress predicted vibration.

[0009] In one possible implementation, during the real-time adaptive motion execution phase, the motion control module employs an adaptive control algorithm based on full-state feedback. The inputs include the planned trajectory from the digital twin and predictive planning module, real-time state feedback from the state perception module, and feedforward control quantities from the digital twin and predictive planning module. The real-time state feedback includes joint angles fed back by the optical encoder, acceleration and angular velocity fed back by the micro IMU, and end effector position deviation fed back by the laser interferometer. The control algorithm integrates this information to dynamically adjust the torque output of each joint motor, ensuring accurate tracking of the planned trajectory and actively suppressing vibration.

[0010] In one possible implementation, during the end-effector alignment and pick-and-place phase, the end effector module performs sub-micron level scanning of the wafer edge or surface features using a linear CCD sensor and a spectral confocal displacement sensor to identify the deviation between the current pose and the desired pose of the wafer. The central coordinating decision-maker generates compensation commands, and the motion control module drives the robotic arm to perform adaptive motion to achieve alignment. At the moment of pick-and-place, the torque sensor is activated, and the control system switches to a compliant control mode to maintain a constant, small contact force and prevent hard contact.

[0011] In one possible implementation, during the health monitoring and particle control phase, the vibration sensor in the state perception module continuously monitors the vibration spectrum of the robotic arm. If an abnormal spectrum is detected, an early warning is issued. At the same time, the micro environmental particle sensor monitors the changes in particle concentration around the robotic arm's movement path. The central collaborative decision-maker records the particle data during the transfer of each wafer and analyzes it in relation to the movement trajectory and speed. If a specific action is found to cause an abnormal increase in particle concentration, the motion parameters of that segment of the trajectory are automatically optimized.

[0012] In one possible implementation, during the task completion and learning optimization phase, the central collaborative decision-maker compares the actual operational data of the task, including the final positioning error, vibration data, and completion time, with the predicted data from the digital twin module. It then uses machine learning algorithms to correct the parameters in the prediction model, thereby enabling the system to continuously self-optimize.

[0013] In one possible implementation, the learning optimization module quantifies the model error by calculating a loss function, which is expressed as:

[0014] in It is the value of the loss function. This is the deviation value for the i-th error type, where i=1,2,3 correspond to the positioning error respectively. Vibration error and time error , It is the weighting coefficient for the i-th error type. It is the difference between the actual end position and the theoretical target position obtained from the state awareness module. It is the difference between the actual vibration amplitude obtained from the vibration sensor and the predicted vibration amplitude. It is the difference between the actual task completion time and the predicted task completion time. It can be preset by the system or adaptively adjusted based on historical data.

[0015] In one possible implementation, the learning optimization module uses a loss function L to calculate the model parameters. Update volume The calculation formula is:

[0016] in Model parameters Update volume It's the learning rate, set by the system. It is a loss function For model parameters The gradient is calculated using numerical difference or backpropagation algorithms.

[0017] In one possible implementation, during particle control optimization, a central collaborative decision-maker automatically adjusts the velocity or acceleration curve of a trajectory segment based on the correlation analysis between particulate matter concentration data and motion parameters to reduce particle generation.

[0018] This specification provides a multimodal adaptive predictive control system for a wafer transfer robotic arm. By integrating multimodal sensing, digital twin simulation, and adaptive control, it achieves high precision and high reliability throughout the wafer transfer process. During system initialization, an automatic calibration model is established to provide an accurate benchmark for subsequent operations. Digital twin simulation predicts vibration and interference in advance, generating optimized trajectories and feedforward control quantities to effectively mitigate operational risks. Real-time adaptive motion integrates multi-source feedback and feedforward compensation to ensure trajectory tracking accuracy and actively suppress vibration. End-effector alignment and compliant control prevent wafer damage and particle generation. Health monitoring and particle control optimize motion parameters in real time to reduce contamination sources. A learning optimization mechanism continuously improves the system through data comparison and model correction, ultimately increasing production efficiency and product yield. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of a multimodal adaptive predictive control system for a wafer transfer robotic arm, provided in one embodiment of this specification. Detailed Implementation

[0020] Many specific details are set forth in the following description to provide a full understanding of this specification. However, this specification can be implemented in many other ways than those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this specification. Therefore, this specification is not limited to the specific implementations disclosed below.

[0021] The terminology used in one or more embodiments of this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the one or more embodiments of this specification. The singular forms “a” and “the” as used in one or more embodiments of this specification and the appended claims are also intended to include the plural forms, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items.

[0022] It should be understood that although the terms first, second, etc., may be used to describe various information in one or more embodiments of this specification, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first may also be referred to as second without departing from the scope of one or more embodiments of this specification, and similarly, second may also be referred to as first. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0023] This specification provides a multimodal adaptive predictive control system for a wafer transfer robot arm, which will be described in detail in the following embodiments.

[0024] See Figure 1 , Figure 1 This document illustrates a system schematic of a multimodal adaptive predictive control system for a wafer transfer robotic arm according to an embodiment of this specification. Specifically, it includes a central collaborative decision-maker, a state perception module, a motion control module, a digital twin and predictive planning module, and an end effector module. After system power-on, the central collaborative decision-maker performs initialization, driving the robotic arm to multiple reference calibration points distributed within the equipment chamber. The state perception module collects actual position data and calibrates kinematic parameters and zero-position deviation to establish a digital model of the robotic arm. Upon receiving transfer tasks from the upper-level management system, the central collaborative decision-maker performs full-path dynamic simulation through the digital twin and predictive planning module, predicting vibration modes and interference risks, and generating the optimal trajectory and feedforward control input. The motion control module performs adaptive motion control based on the optimal trajectory and feedforward control input, combined with real-time state feedback from the state perception module. The end effector module scans the wafer pose using sensors for alignment and placement. The state perception module continuously monitors the vibration spectrum and particulate matter concentration. The central collaborative decision-maker records operational data and compares it with predicted data, correcting model parameters through a learning optimization algorithm.

[0025] The system comprises several key components: a central coordinating decision-maker (CCD) – the core control unit that coordinates the operation and decision-making processes of all modules, and performs initialization and task planning; a state awareness module – a data acquisition module integrating vision, laser, vibration, and force sensors, providing real-time status information on the robotic arm's position, vibration, and environment; a motion control module – a motion drive module employing adaptive algorithms, precisely controlling the torque and trajectory of the robotic arm's joint motors; a digital twin and predictive planning module – a virtual simulation module based on a digital model, simulating the robotic arm's dynamic behavior and predicting risks in a virtual space; an end effector module – a precision control module dedicated to wafer alignment and placement, using high-resolution sensors to achieve sub-micron level pose adjustment; system power-on – the power connection process that triggers system initialization and self-calibration; initialization – a series of preset calibration and setup steps that drive the robotic arm to a reference point and establish a digital model; and a robotic arm – a multi-axis motion device that performs wafer transfer for precise positional movement within the cavity. The equipment chamber refers to the enclosed space housing the robotic arm and wafer processing equipment, providing a fixed environment for reference calibration points. Reference calibration points are reference points with known locations within the chamber, enabling model calibration through sensor measurements. Actual position data refers to the true coordinate information of the robotic arm's end effector acquired by sensors, used to compare with the theoretical model to calculate deviations. Kinematic parameters (such as DH parameters) are mathematical parameters describing the robotic arm's motion relationships, adjusted during calibration to improve model accuracy. Zero-position deviation refers to the error values ​​of the zero positions of each axis of the robotic arm, used to eliminate initial position offsets through calibration. The digital model refers to a virtual representation of the robotic arm built based on calibration data, used for simulation and predictive control. The upper-level management system refers to an external control or production management system that can issue wafer transfer task instructions. A transfer task refers to the wafer movement operation from the pick-up position to the place-down position, specifying theoretical coordinates and path requirements. Full-path dynamic simulation refers to simulating the entire motion process of the robotic arm in a digital twin environment, used to predict vibration and interference risks. Vibration modes refer to the structural vibration characteristics (such as frequency and amplitude) excited by the robotic arm during motion, which can be identified in advance through simulation. Interference risk refers to the possibility of the robotic arm colliding with surrounding equipment, which can be used to optimize the trajectory to avoid damage. Optimal trajectory refers to a smooth, collision-free motion path verified by simulation, used as a setpoint for real-time control. Feedforward control quantity refers to a compensation signal calculated based on predicted vibration, used to actively suppress vibration in control. Real-time state feedback refers to the robotic arm state data (such as joint angles and acceleration) acquired in real time from sensors, which can be used for adaptive control adjustments. Adaptive motion control refers to a control mode that dynamically adjusts torque based on feedback and feedforward, ensuring trajectory tracking accuracy. Sensors refer to detection devices on the end effector (such as CCDs or spectral confocal displacement sensors) to scan wafer features and acquire pose data.Wafer pose refers to the position and orientation information of a wafer in space, used to calculate minute deviations from the desired pose. Alignment refers to the minute position adjustment process at the end effector of a robotic arm, enabling precise alignment of the wafer with the actuator. Pick-and-place refers to the control mode of picking up or placing wafers with a constant, minute force, used to prevent hard contact and damage. Vibration spectrum refers to the frequency distribution characteristics of the robotic arm's vibration signal, used for health monitoring and anomaly detection. Particulate matter concentration refers to the level of particles in the air surrounding the robotic arm, which can be correlated with motion parameters to optimize control. Operational data refers to actual performance records during task execution (such as errors and time), used for comparative analysis with predicted data. Predicted data refers to the expected results generated by digital twin simulation (such as vibration and trajectory), used as an optimization benchmark. Learning optimization algorithms refer to parameter correction methods based on machine learning, which can improve model accuracy through error analysis. Model parameters refer to adjustable variables in digital twins and control algorithms, used to improve system performance in iterations.

[0026] As a concrete example: After the system is powered on, the central collaborative decision-maker initiates initialization, driving the robotic arm to move at low speed to three reference calibration points within the equipment chamber (e.g., coordinates X=100mm, Y=200mm, Z=50mm). The vision sensor and laser tracker in the state perception module collect actual position data, compare it with the theoretical model, and automatically calibrate kinematic parameters (such as link length and joint angles in the DH parameters) and zero-point deviation (e.g., adjusting the zero-point offset of each axis to less than 0.01mm), establishing a digital model. Subsequently, the upper-level management system issues a transfer task (retrieving the wafer from the transfer chamber to the process chamber, wafer retrieval position coordinates A (150mm, 250mm, 0mm), wafer placement position coordinates B (300mm, 400mm, ...). The digital twin and predictive planning module performs full-path dynamic simulation, predicting that the robotic arm will generate a 10Hz vibration mode and interference risk with surrounding equipment at a speed of V=0.5m / s. It generates an optimal trajectory (e.g., a Bézier curve path) and feedforward control inputs (e.g., compensating torque of 0.1Nm). Based on the optimal trajectory and feedforward control inputs, and combined with real-time status feedback from the state perception module (e.g., optical encoder feedback joint angle error <0.001rad, IMU feedback peak acceleration <2g, laser interferometer feedback end-effector position deviation <1μm), the motion control module performs adaptive motion control, dynamically adjusting the motor current of each joint to suppress vibration. When the robotic arm's end-effector approaches the wafer, the end effector module scans the wafer edge using a linear CCD sensor and a spectral confocal displacement sensor, identifying the wafer pose and its relationship to the desired value in the X and Y axes. For minor deviations in the Z and θz directions (e.g., X deviation -0.5μm, Y deviation +0.3μm), the central collaborative decision-maker generates compensation commands to drive the robotic arm to perform adaptive motion for alignment. At the moment of picking up and placing the piece, the torque sensor is activated, and the control system switches to compliant control mode to maintain the contact force constant below 0.1N to complete the picking and placing. Throughout the process, the state perception module continuously monitors the vibration spectrum (e.g., baseline spectrum 50-100Hz, with an abnormal peak of 150Hz) and particulate matter concentration (e.g., normal value <10 particles / m³, abnormal value >50 particles / m³). The central collaborative decision-maker records the running data (e.g., actual positioning error 0.2μm, vibration amplitude 0.05g, completion time 2.5s) and compares it with the predicted data. By learning optimization algorithms (e.g., gradient descent method), the model parameters are corrected (e.g., adjusting DH parameters or control gain) to achieve continuous system optimization.

[0027] This invention improves the accuracy and reliability of wafer transfer through multi-module collaboration and adaptive predictive control. Automatic calibration during system initialization ensures the accuracy of the digital model, digital twin simulation avoids motion risks and vibrations in advance, real-time adaptive control integrates multi-source feedback to achieve high-precision trajectory tracking, end-point alignment and placement prevent wafer damage and particle generation, health monitoring optimizes motion parameters in real time to reduce contamination, and the learning optimization mechanism continuously improves system performance through data comparison, ultimately enhancing production efficiency and product yield.

[0028] In one possible implementation, during the initialization phase, the state perception module includes a vision sensor and a laser tracker to sample and measure the actual position of the robotic arm's end effector and compare it with the theoretical model to automatically calibrate the robotic arm's kinematic parameters and zero-position deviation. Simultaneously, a vibration sensor collects background noise and establishes benchmark reference data for vibration monitoring based on the collected background noise, while a force sensor performs zero-drift calibration.

[0029] Among these, a visual sensor can refer to a position detection device based on optical imaging, capable of acquiring two-dimensional or three-dimensional spatial coordinate data of the robotic arm's end effector. A laser tracker can refer to a precision measuring instrument utilizing the principle of laser interferometry, used to obtain high-precision three-dimensional position information of the robotic arm's end effector in space. A theoretical model can refer to the ideal kinematic equations established based on the robotic arm's design parameters, serving as a benchmark for comparison with actual data during calibration. Kinematic parameters can refer to mathematical variables describing the length of the robotic arm's links, joint angles, and geometric relationships between axes, used to correct the accuracy of the robotic arm's digital model during calibration. Zero-point deviation can refer to the angular or positional difference between the actual mechanical zero point and the theoretical zero point of each joint axis of the robotic arm, which can be compensated for through calibration to improve initial positioning accuracy. A vibration sensor can refer to a sensing device that detects vibration signals of a mechanical structure, used to collect the background noise spectrum characteristics of the equipment's operating environment. Background noise acquisition refers to the process of recording environmental vibration signals in a motionless state, enabling the establishment of benchmark reference data for vibration monitoring. A force sensor can refer to a sensing element that measures contact force and torque, capable of detecting the sensor's own zero-point output drift. Zero-drift calibration refers to adjusting the output signal of a force sensor to zero under no-load conditions to eliminate inherent biases in the measurement system.

[0030] As a concrete example: During the initialization phase, the state perception module activates a vision sensor (such as an industrial camera with a resolution of 2048×2048) and a laser tracker (such as a laser interferometer system with a measurement accuracy of 0.1μm) to sample and measure the end effector of the robotic arm that has moved to the reference calibration point (such as coordinates X=100.000mm, Y=200.000mm, Z=50.000mm) to obtain the actual position data (such as measured values ​​X=100.002mm, Y=199.998mm). Z=50.001mm); The actual position data is compared with the theoretical model (such as the homogeneous transformation matrix established based on DH parameters), and the kinematic parameters (such as adjusting the link length parameter from the theoretical value of 150.000mm to 150.001mm) and zero-position deviation (such as compensating for the zero-position offset angle of joint 1 of -0.005°) are automatically calibrated. At the same time, the vibration sensor (such as the accelerometer with a frequency range of 0.5-2000Hz) collects background noise when the robot arm is stationary (recording the peak values ​​of the spectrum at 60Hz and 120Hz), and the force sensor (such as the six-dimensional force sensor with a range of ±10N) performs zero drift calibration (adjusting the initial output value from 0.05N to 0.00N).

[0031] This invention improves the initial accuracy of the system through multi-sensor collaborative calibration. The vision sensor and laser tracker provide high-precision position measurement data, which are compared with the theoretical model to automatically correct kinematic parameters and zero-position deviation, ensuring that the digital model is consistent with the actual robotic arm. The vibration sensor collects background noise to establish a vibration monitoring baseline, and the force sensor zero-drift calibration eliminates measurement system errors, laying the foundation for subsequent precise control and health monitoring.

[0032] In one possible implementation, during the task planning and digital twin simulation phase, the digital twin and predictive planning module calls the calibrated digital model of the robotic arm to perform full-path dynamic simulation in virtual space. Based on the reference data, it calculates the vibration modes, motion time, and interference risk of the robotic arm at each velocity point, generates a smooth and collision-free optimal trajectory, and calculates the feedforward control quantity required to suppress predicted vibration.

[0033] The digital twin and predictive planning module refers to a virtual simulation system based on a digital model, capable of calling upon a calibrated robotic arm digital model for dynamic simulation. The calibrated robotic arm digital model refers to a high-precision virtual model with parameter corrections during the initialization phase, used to accurately reflect the motion characteristics of the actual robotic arm. Virtual space refers to a computer-generated 3D simulation environment capable of simulating the robotic arm's motion within a real equipment chamber. Full-path dynamic simulation refers to simulating the complete motion process of the robotic arm from start to finish, used to predict various dynamic characteristics during the motion. Vibration modes refer to the inherent vibration patterns of the robotic arm structure excited at specific velocity points, enabling the early identification of potential vibration risks through simulation. Motion time refers to the duration required for the robotic arm to complete a specified trajectory, used to evaluate the efficiency of the transfer task. Interference risk refers to the possibility of the robotic arm colliding with surrounding equipment during motion, which can be identified and avoided in advance. Optimal trajectory refers to a smooth motion path optimized through simulation, used to achieve a collision-free and vibration-minimized transfer process. Feedforward control refers to a compensation control signal based on predicted vibration calculations, capable of applying control in advance to suppress expected vibrations.

[0034] As a concrete example: the digital twin and predictive planning module calls the calibrated digital model of the robotic arm (including corrected DH parameters and zero-point compensation) and performs full-path dynamic simulation in virtual space (simulating the actual equipment chamber size of 2000mm×2000mm×1500mm); based on the benchmark reference data obtained in the above scheme, the simulation calculates the vibration modes of the robotic arm at velocity points of 0.3m / s, 0.5m / s and 0.7m / s (identifying resonance risks at 12Hz and 25Hz), predicts the motion time to be 3.2 seconds, and detects the interference risk of only 2mm distance between the robotic arm and the process chamber door frame; based on the simulation results, a smooth and collision-free optimal trajectory is generated (using a fifth-order B-spline curve to connect 24 path points), and the feedforward control quantity (including the compensation torque of 0.05-0.15Nm for each joint during the acceleration phase) is calculated.

[0035] This invention achieves pre-verification and optimization of the motion process through digital twin simulation. It uses a calibrated and accurate model to perform full-path dynamic simulation in virtual space, accurately predicts the vibration modes and interference risks at each velocity point, generates a smooth and collision-free optimal trajectory, and calculates feedforward control quantities. This effectively avoids collision accidents and vibration problems in actual operation, and improves the safety and stability of system operation.

[0036] In one possible implementation, during the real-time adaptive motion execution phase, the motion control module employs an adaptive control algorithm based on full-state feedback. The inputs include the planned trajectory from the digital twin and predictive planning module, real-time state feedback from the state perception module, and feedforward control quantities from the digital twin and predictive planning module. The real-time state feedback includes joint angles fed back by the optical encoder, acceleration and angular velocity fed back by the micro IMU, and end effector position deviation fed back by the laser interferometer. The control algorithm integrates this information to dynamically adjust the torque output of each joint motor, ensuring accurate tracking of the planned trajectory and actively suppressing vibration.

[0037] Among these, the full-state feedback adaptive control algorithm refers to a control method that dynamically adjusts multiple real-time measurement signals, automatically adjusting control parameters according to changes in system state. The planned trajectory refers to the ideal motion path coordinate sequence generated by the digital twin module, used as a tracking target for real-time motion control. Real-time state feedback refers to the robotic arm motion parameters continuously acquired from multi-dimensional sensing modules, providing actual motion information for joints and the end effector. Feedforward control refers to the control compensation signal pre-calculated based on vibration prediction, used for active suppression before vibration occurs. An optical encoder refers to a high-precision angle measurement device installed at the joint, providing real-time data on joint angles. Joint angles refer to the actual rotational positions of each joint axis of the robotic arm, used to calculate deviations from the planned trajectory. A miniature IMU refers to a small inertial measurement unit integrated into each joint of the robotic arm, capable of measuring the linear acceleration and angular velocity of the joint. Acceleration refers to the linear acceleration value during joint movement, reflecting the dynamic characteristics of the robotic arm. Angular velocity refers to the measured angular velocity of joint rotation, used to detect abnormal vibration states. A laser interferometer can refer to a precision position measuring instrument based on the principle of laser interferometry, used to detect the spatial position deviation of an end effector. End effector position deviation refers to the minute difference between the actual and theoretical positions of the end effector, which can serve as a basis for control compensation. Torque output refers to the torque command calculated by the control algorithm and sent to the joint motors, used to precisely drive the movement of the robotic arm.

[0038] As a concrete example: The motion control module runs an adaptive control algorithm based on full-state feedback, receiving the planned trajectory (containing the coordinate sequence of 500 path points) from the digital twin module, real-time state feedback from the state perception module (optical encoder feedback joint angle error of 0.002 rad, micro-IMU feedback peak acceleration of 1.8 m / s² and angular velocity of 0.5 rad / s, laser interferometer feedback end-effector position deviation of 0.8 μm), and feedforward control input from the digital twin module (compensation torque of 0.12 Nm for joint 2). After fusing this information, the control algorithm dynamically adjusts the torque output of each joint motor (adjusting the torque of joint 1 from 0.25 Nm to 0.28 Nm and the torque of joint 3 from 0.18 Nm to 0.16 Nm), so that the tracking error between the actual motion trajectory of the robotic arm end and the planned trajectory is less than 5 μm, and suppressing the expected 12 Hz vibration amplitude at the 0.5 m / s velocity point to less than 20% of the original value.

[0039] This invention achieves high-precision motion tracking through full-state feedback and adaptive control. It integrates planned trajectory, multi-source real-time feedback and feedforward compensation, and dynamically adjusts the torque output of each joint to ensure that the robotic arm accurately tracks the predetermined trajectory. At the same time, it actively suppresses structural vibration and improves motion stability and positioning accuracy.

[0040] In one possible implementation, during the end-effector alignment and pick-and-place phase, the end effector module performs sub-micron level scanning of the wafer edge or surface features using a linear CCD sensor and a spectral confocal displacement sensor to identify the deviation between the current pose and the desired pose of the wafer. The central coordinating decision-maker generates compensation commands, and the motion control module drives the robotic arm to perform adaptive motion to achieve alignment. At the moment of pick-and-place, the torque sensor is activated, and the control system switches to a compliant control mode to maintain a constant, small contact force and prevent hard contact.

[0041] Among these, a linear CCD sensor refers to an optical imaging device employing a linear charge-coupled element (LCA) for continuous scanning and feature extraction of the wafer edge contour. A spectral confocal displacement sensor refers to a non-contact ranging device based on the principle of spectral confocality, capable of accurately measuring the relative distance between the wafer surface and the sensor. Submicron-level scanning refers to a precision measurement process with submicron-level resolution, used to acquire subtle feature information of the wafer edge or surface. A wafer edge refers to the annular region around the outer circumference of the wafer, providing positioning reference features. Surface features refer to specific markings or textures on the wafer's surface, aiding in pose recognition. The current pose refers to the wafer's actual position and orientation angle in space, used for comparison with the desired state. The desired pose refers to the ideal spatial state of the wafer at the target position, serving as a benchmark for alignment operations. Micro-deviation refers to the minute difference between the current pose and the desired pose, used as the basis for generating compensation commands. Compensation commands refer to position correction commands calculated by a central collaborative decision-maker, guiding the robotic arm to adjust its pose. Micro-adaptive motion refers to millimeter- or micrometer-level precision movements performed by the end effector of a robotic arm to achieve accurate wafer alignment. Torque sensors are devices that detect contact forces and torques to monitor the contact state during pick-and-place operations. Compliant control modes refer to control strategies that automatically adjust position based on force feedback, maintaining stable contact force output. Constant micro-contact force refers to a stable and minute force value maintained during pick-and-place operations to prevent hard contact between the wafer and the carrier. Hard contact refers to a rigid collision between the wafer and the contact surface, which may cause wafer damage or particle generation.

[0042] As a concrete example: the end effector module activates a linear CCD sensor (0.5μm resolution) and a spectral confocal displacement sensor (0.1μm measurement accuracy) to perform sub-micron scanning of the edge of a 300mm wafer, acquiring contour data from 256 sampling points; through feature matching, it identifies the micro-deviations between the wafer's current pose (position deviation X=-0.8μm, Y=+0.6μm, angular deviation θz=0.02°) and the desired pose; the central collaborative decision-maker generates compensation commands (including X-axis compensation). The Y-axis compensation is +0.8μm, the Y-axis compensation is -0.6μm, and the θz-axis compensation is -0.02°. The motion control module drives the robotic arm to perform adaptive motion (movement range ±10μm, speed 0.1mm / s) to achieve precise alignment. At the moment of wafer pick-up, the torque sensor (range ±5N) is activated and detects that the contact force has risen to 0.08N. The control system immediately switches to compliant control mode and maintains the contact force constant within the range of 0.1N±0.02N through PID adjustment, completing the wafer pick-up and placement operation without impact.

[0043] This invention achieves safe and precise wafer operation through multi-sensor fusion and compliant control. High-resolution sensors provide sub-micron level scanning data, accurately identify micro-deviations in position and generate compensation commands, and achieve precise alignment through minute adaptive motion. Force feedback triggers a compliant control mode to maintain a constant minute contact force, effectively avoiding wafer damage and particle contamination caused by hard contact, thereby improving operational safety and yield.

[0044] In one possible implementation, during the health monitoring and particle control phase, the vibration sensor in the state perception module continuously monitors the vibration spectrum of the robotic arm. If an abnormal spectrum is detected, an early warning is issued. At the same time, the micro environmental particle sensor monitors the changes in particle concentration around the robotic arm's movement path. The central collaborative decision-maker records the particle data during the transfer of each wafer and analyzes it in relation to the movement trajectory and speed. If a specific action is found to cause an abnormal increase in particle concentration, the motion parameters of that segment of the trajectory are automatically optimized.

[0045] Among these, abnormal spectrum refers to frequency components in vibration signals that deviate from normal patterns, indicating potential structural abnormalities or component wear in the robotic arm. Early warning refers to alarm signals issued when the system detects an anomaly, prompting the need for equipment maintenance or inspection. Miniature environmental particle sensors are devices that monitor the concentration of airborne particles, used to measure particulate matter levels in the robotic arm's movement area in real time. Particulate matter concentration changes refer to fluctuations in the number of particles per unit volume of air, reflecting the impact of robotic arm movement on the clean environment. Particulate matter data refers to the sequence of concentration measurements collected by particle sensors, used for correlation analysis with motion parameters. Motion trajectory refers to the actual movement path record of the robotic arm in space, serving as a basis for analyzing the sources of particulate matter generation. Velocity correlation analysis involves analyzing the correlation between particulate matter data and robotic arm movement speed data to identify motion patterns leading to increased particulate matter levels. Specific actions refer to a specific movement segment or operation sequence of the robotic arm, which can be analyzed to determine its causal relationship with particulate matter generation. Abnormal increase in particulate matter concentration refers to the phenomenon where particulate matter concentration exceeds a preset threshold, triggering the system's automatic optimization process. Motion parameters can refer to variables such as acceleration and velocity curves that control the movement of a robotic arm, and can be adjusted to reduce particulate matter generation.

[0046] As a specific example: During the transfer of the 25th wafer, the vibration sensor detected abnormal frequencies at 85Hz and 170Hz (amplitude exceeding the baseline by 3dB), and the system immediately issued an early warning to the monitoring terminal; at the same time, the micro environmental particle sensor detected that the particle concentration suddenly increased from 5 particles / m³ to 35 particles / m³ when the robotic arm passed the path point P (350mm, 280mm); the central collaborative decision-maker recorded the particle data during the transfer of this wafer (peak concentration 35 particles / m³, duration 0.8s) and performed correlation analysis with the motion trajectory (path segment S12) and speed (acceleration segment 0.3-0.7m / s²), confirming that the acceleration action caused the abnormal increase in particle concentration; the system automatically optimized the motion parameters of this trajectory segment, reducing the maximum acceleration from 0.7m / s² to 0.4m / s², and after optimization, the peak particle concentration of the same path segment decreased to 12 particles / m³.

[0047] This invention enables proactive health management by real-time monitoring of vibration spectrum and particulate matter concentration. Vibration anomaly warning can detect potential equipment failures in advance, and the correlation analysis of particulate matter monitoring and motion parameters can accurately locate pollution sources. Automatic optimization of motion parameters reduces particulate matter generation from the source, effectively maintaining a clean environment and extending equipment lifespan.

[0048] In one possible implementation, during the task completion and learning optimization phase, the central collaborative decision-maker compares the actual operational data of the task, including the final positioning error, vibration data, and completion time, with the predicted data from the digital twin module. It then uses machine learning algorithms to correct the parameters in the prediction model, thereby enabling the system to continuously self-optimize.

[0049] Among these, actual operational data refers to the set of real performance indicators collected during task execution, reflecting the actual working state of the robotic arm. Final positioning error refers to the spatial deviation between the robotic arm's end effector and the target position upon task completion, used to assess positioning accuracy. Vibration data refers to the characteristic values ​​of vibration signals generated by the robotic arm during movement, reflecting the dynamic performance of the mechanical structure. Completion time refers to the actual time elapsed from the start to the end of the task, used to assess work efficiency. Predictive data refers to the theoretical performance indicators pre-calculated by the digital twin module, serving as a benchmark for comparative analysis. Machine learning algorithms refer to computational methods that automatically improve model performance through data training, used to correct parameter biases in the predictive model. The predictive model refers to the digital twin mathematical model used for simulation and performance prediction, which can improve prediction accuracy through parameter adjustment. Parameters refer to the adjustable coefficient variables in the predictive model, which can continuously improve model performance through optimization. Continuous system self-optimization refers to the process of continuously improving system performance through iterative learning, enabling long-term performance improvement.

[0050] As a specific example: After the 50th wafer transfer task was completed, the central collaborative decision-maker collected actual operational data, including the final positioning error of 0.3 μm, vibration data (main frequency component 15 Hz amplitude 0.02 g), and completion time of 2.8 seconds, and compared them with the predicted data from the digital twin module (positioning error 0.1 μm, vibration frequency 18 Hz amplitude 0.015 g, completion time 2.5 seconds). The machine learning algorithm (using random forest regression) was used to calculate the parameter correction in the prediction model, adjusting the link stiffness parameter in the kinematic model from the theoretical value of 5.0 × 10^6 N / m to 5.2 × 10^6 N / m, and the damping coefficient from 150 N·s / m to 145 N·s / m. After parameter correction, the error between the predicted and actual data was reduced by approximately 40% in the 51st task, achieving continuous self-optimization of the system.

[0051] This invention uses machine learning algorithms to automatically correct the parameters of the prediction model by comparing and analyzing actual operating data with predicted data. This enables the system to continuously adapt to changes in equipment status, continuously improve prediction accuracy and control performance, and achieve long-term stable operation and performance self-improvement.

[0052] In one possible implementation, the learning optimization module quantifies the model error by calculating a loss function, which is expressed as:

[0053] in It is the value of the loss function. This is the deviation value for the i-th error type, where i=1,2,3 correspond to the positioning error respectively. Vibration error and time error , It is the weighting coefficient for the i-th error type. It is the difference between the actual end position and the theoretical target position obtained from the state awareness module. It is the difference between the actual vibration amplitude obtained from the vibration sensor and the predicted vibration amplitude. It is the difference between the actual task completion time and the predicted task completion time. It can be preset by the system or adaptively adjusted based on historical data.

[0054] The learning optimization module refers to the processing unit that calculates the loss function using an algorithm to optimize the model, quantifying model errors and adjusting prediction model parameters. The loss function value refers to the numerical result obtained from the loss function calculation, reflecting the overall deviation between the model's prediction and the actual data. The error type deviation value refers to the difference between the measured and predicted values ​​for a specific error category, used to calculate the contribution of each error in the loss function. The weighting coefficient refers to the proportional factors assigned to different error types in the loss function, adjusting the importance of each error term in the total loss.

[0055] As a concrete example: After completing the 100th wafer transfer task, the learning optimization module obtains the actual end position (measured values ​​X=300.002mm, Y=400.001mm, Z=0.000mm) and the theoretical target position (X=300.000mm, Y=400.000mm, Z=0.000mm) from the state awareness module and calculates the positioning error. =0.002mm, calculate the vibration error by comparing the actual vibration amplitude of 0.03g obtained from the vibration sensor with the predicted vibration amplitude of 0.02g. =0.01g, the actual task completion time of 2.9 seconds obtained from the system record and the predicted task completion time of 2.7 seconds are used to calculate the time error. =0.2 seconds; Set weighting coefficient =0.5 (positioning error weight) =0.3 (vibration error weight) =0.2 (time error weight), calculate the loss function value L=0.5×(0.002)^2 + 0.3×(0.01)^2 + 0.2×(0.2)^2 = 0.000002 + 0.00003 + 0.008 =0.008032; based on the loss function value, adjust the stiffness and damping parameters in the prediction model using the gradient descent algorithm to reduce the loss function value to below 0.005 in the next task.

[0056] This invention accurately quantifies model errors through a loss function, comprehensively evaluates system performance by integrating the weighted sum of squared errors in positioning, vibration, and time, flexibly adjusts optimization priorities using weight coefficients, and continuously reduces loss values ​​through iterative calculations, effectively improving the accuracy of the prediction model and control precision, and achieving a stable improvement in system performance.

[0057] In one possible implementation, the learning optimization module uses a loss function L to calculate the model parameters. Update volume The calculation formula is:

[0058] in Model parameters Update volume It's the learning rate, set by the system. It is a loss function For model parameters The gradient is calculated using numerical difference or backpropagation algorithms.

[0059] In this context, model parameters θ refer to the set of variables that need to be adjusted in the prediction model, controlling the model's behavior and accuracy. The update increment Δθ refers to the amount of adjustment to the model parameters in each iteration, enabling the model to predict more closely to the actual data. The learning rate α is a hyperparameter that controls the step size of parameter updates, affecting the convergence speed of the optimization process. Gradient L / θ can refer to the rate of change of the loss function with respect to the model parameters, indicating the direction and magnitude of parameter adjustment. Numerical differencing refers to a numerical method that approximates the gradient by calculating the difference between function values, used when analytical gradients are unavailable. Backpropagation refers to an algorithm based on the chain rule for calculating gradients, capable of efficiently computing gradients in neural networks.

[0060] As a concrete example: After completing the 200th wafer transfer task, the learning optimization module, based on the loss function L=0.005 (where the positioning error...),... =0.002mm, vibration error =0.01g, time error =0.2 seconds, weighting coefficient =0.5、 =0.3、 =0.2), set the learning rate α=0.01; calculate the gradient using the numerical difference method. The gradient value is -0.15; then the update amount Δθ = -0.01 × (-0.15) = 0.0015 is calculated; the model parameter θ is adjusted from the current value of 1.2000 to 1.2015; after the parameter update, the loss function value is reduced to 0.0038 in the next task, and the model prediction accuracy is significantly improved.

[0061] This invention optimizes the model through gradient calculation and parameter updates. It uses the learning rate to control the step size to avoid over-updating, the gradient indicates the descent direction of the loss function, and numerical differencing or backpropagation provides flexibility in gradient calculation, ensuring continuous optimization of model parameters and improving the system's prediction accuracy and adaptability.

[0062] In one possible implementation, during particle control optimization, a central collaborative decision-maker automatically adjusts the velocity or acceleration curve of a trajectory segment based on the correlation analysis between particulate matter concentration data and motion parameters to reduce particle generation.

[0063] Among these, particulate matter concentration data refers to the sequence of measurements of the number of airborne particles collected from particle sensors, reflecting the impact of robotic arm movement on environmental cleanliness. Motion parameters refer to the set of physical quantities controlling the robotic arm's movement, including variables such as velocity, acceleration, and jerk. Correlation analysis refers to the method of calculating the correlation between particulate matter data and motion parameters, which can identify key motion factors leading to particulate matter generation. Trajectory segments refer to specific intervals within the robotic arm's motion path, serving as optimization units for particulate matter control. Velocity curves refer to the functional relationship between the robotic arm's velocity and time during movement, which can be adjusted to reduce particle generation. Acceleration curves refer to the functional relationship between the robotic arm's acceleration and time, which can be optimized to reduce particulate matter caused by motion impact.

[0064] As a specific example: When the central collaborative decision-maker analyzed the data from the 80th wafer transfer mission, it found that when the robotic arm passed through trajectory segment T5 (a straight path from point P3 to P4), the particulate matter concentration suddenly increased from the baseline value of 8 particles / m³ to 42 particles / m³. At the same time, the motion parameters of this trajectory segment were recorded as velocity 0.6 m / s and acceleration 0.5 m / s². Through correlation analysis, it was determined that the correlation coefficient between the peak acceleration and the increase in particulate matter concentration reached 0.85. The system automatically adjusted the acceleration curve of this trajectory segment from the original 0.5 m / s² to 0.3 m / s², and the velocity curve from 0.6 m / s to 0.5 m / s. After optimization, the peak particulate matter concentration of the same trajectory segment decreased to 18 particles / m³, while maintaining the positioning accuracy within the allowable range.

[0065] This invention achieves precise optimization through correlation analysis of particulate matter data and motion parameters. It automatically identifies motion characteristics that lead to increased particulate matter levels and adjusts the velocity and acceleration curves of the trajectory segment accordingly. While ensuring motion accuracy, it effectively reduces particulate matter generation, maintains the quality of the cleanroom environment, and reduces the risk of wafer contamination.

[0066] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments in this specification are not limited to the described order of actions, because according to the embodiments in this specification, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the embodiments in this specification.

[0067] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0068] The preferred embodiments disclosed above are merely illustrative of this specification. The optional embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the embodiments described herein. These embodiments are selected and specifically described in this specification to better explain the principles and practical applications of the embodiments, thereby enabling those skilled in the art to better understand and utilize this specification. This specification is limited only by the claims and their full scope and equivalents.

Claims

1. A multimodal adaptive predictive control system for a wafer transfer robotic arm, characterized in that, The system includes a central collaborative decision-maker, a state perception module, a motion control module, a digital twin and predictive planning module, and an end effector module. After the system is powered on, the central collaborative decision-maker performs initialization, drives the robotic arm to move to multiple reference calibration points distributed in the equipment chamber, collects actual position data through the state perception module and calibrates kinematic parameters and zero-position deviation, and establishes a digital model of the robotic arm. After receiving the task transmitted by the upper management system, the central collaborative decision-maker performs full-path dynamic simulation through the digital twin and predictive planning module to predict vibration modes and interference risks, and generate the optimal trajectory and feedforward control quantity. The motion control module performs adaptive motion control based on the optimal trajectory and the feedforward control quantity, combined with the real-time status feedback from the state perception module. The end effector module uses sensors to scan the wafer pose for alignment and placement. The state perception module continuously monitors the vibration spectrum and particulate matter concentration, and the central collaborative decision-maker records the operating data and compares it with the predicted data, and corrects the model parameters through learning optimization algorithms.

2. The system according to claim 1, characterized in that, During the initialization phase, the state perception module includes a vision sensor and a laser tracker, which are used to sample and measure the actual position of the robotic arm's end effector and compare it with the theoretical model to automatically calibrate the robotic arm's kinematic parameters and zero-position deviation. It also includes: a vibration sensor to collect background noise, establishing reference data for vibration monitoring based on the collected background noise, and a force sensor to perform zero drift calibration.

3. The system according to claim 2, characterized in that, During the task planning and digital twin simulation phase, the digital twin and predictive planning module calls the calibrated digital model of the robotic arm to perform full-path dynamic simulation in virtual space. Based on the benchmark reference data, it calculates the vibration mode, motion time and interference risk of the robotic arm at each speed point, generates the optimal trajectory, and calculates the feedforward control quantity required to suppress predicted vibration.

4. The system according to claim 1, characterized in that, During the real-time adaptive motion execution phase, the motion control module employs an adaptive control algorithm based on full-state feedback. The inputs include the planned trajectory from the digital twin and predictive planning module, the real-time state feedback from the state perception module, and the feedforward control quantity from the digital twin and predictive planning module. The real-time state feedback includes the joint angles fed back by the optical encoder, the acceleration and angular velocity fed back by the micro IMU, and the end effector position deviation fed back by the laser interferometer. The control algorithm integrates this information to dynamically adjust the torque output of each joint motor.

5. The system according to claim 1, characterized in that, During the end alignment and pick-and-place phase, the end effector module scans the wafer edge or surface features using a linear CCD sensor and a spectral confocal displacement sensor to identify the deviation between the wafer's current pose and the desired pose. The central collaborative decision-maker generates compensation commands, and the motion control module drives the robotic arm to perform adaptive motion to achieve alignment. When picking and placing wafers, the torque sensor is activated, and the control system switches to compliant control mode.

6. The system according to claim 1, characterized in that, During the health monitoring and particle control phase, the vibration sensor in the state perception module continuously monitors the vibration spectrum of the robotic arm. If an abnormal spectrum is detected, an early warning is issued. The micro environmental particle sensor monitors the changes in particle concentration around the robotic arm's movement path. The central collaborative decision-maker records the particle data during the transfer of each wafer and analyzes it in relation to the movement trajectory and speed. If a specific action is found to cause an abnormal increase in particle concentration, the motion parameters of that segment of the trajectory are automatically optimized.

7. The system according to claim 1, characterized in that, During the task completion and learning optimization phase, the central collaborative decision-maker compares the actual operation data of this task, including the final positioning error, vibration data, and completion time, with the predicted data of the digital twin module, and corrects the parameters in the prediction model through machine learning algorithms.

8. The system according to claim 7, characterized in that, The learning optimization module quantifies model error by calculating a loss function, the formula of which is: in It is the value of the loss function. This is the deviation value for the i-th error type, where i=1,2,3 correspond to the positioning error respectively. Vibration error and time error , It is the weighting coefficient for the i-th error type.

9. The system according to claim 8, characterized in that, The learning optimization module uses the loss function L to calculate model parameters. Update volume The calculation formula is: in Model parameters Update volume It's the learning rate. It is a loss function For model parameters The gradient.

10. The system according to claim 6, characterized in that, In particle control optimization, the central collaborative decision-maker automatically adjusts the velocity or acceleration curve of the trajectory segment based on the correlation analysis between particle concentration data and motion parameters.

Images