Review on Vibration Control of Wafer Handling Robot

Wafer handling robots have been widely used to process silicon wafers in semiconductor manufacturing machines. However, their accuracy, speed and versatility are limited by vibration, which varies with their position in the workspace. One of the main challenges of controlling wafer processing robots is to avoid vibration as much as possible during high-speed motion, which is the main cause of wafer sliding and particle pollution. In order to improve stiffness and reduce vibration, the robot adopts parallel structure, but it cannot meet the requirements of high-precision control. The control of the whole system is affected by many factors, among which the control algorithm, assembly accuracy and machining accuracy of each component are the main influencing factors. Moreover, the vibration caused by structural factors is difficult to detect by the motor encoder and control by the motor. According to the current processing technology, the precision of components is already very high, and it needs to pay a lot of economic costs to further improve the precision of components. It is not appropriate to comprehensively consider the perspective of improving the precision of components, but optimizing trajectory planning and optimizing control algorithms is a better idea. 3.1. Trajectory Planning Research Robot trajectory planning is very important for robot motion control system, which directly affects the performance of robot, which is reflected in many aspects such as operation efficiency, motion stability, energy consumption, etc. For specific robot trajectory planning problems, it is necessary to comprehensively consider the actual needs. Robot cluster tools used in semiconductor manufacturing have strict time constraints, and robot motion should not only consider efficiency, but also run smoothly, which requires smooth trajectories without sudden changes in acceleration [31,32], as shown in Figure 8A. Robot motion planning is also one of the most important problems in robot control. If time optimality and stationarity are considered synthetically, trajectory planning is a smooth time optimal trajectory planning problem. Most existing online trajectory generators aim to achieve good tracking performance by limiting the bandwidth of the reference input. This strategy is unacceptable when applied to wafer handling robots pursuing high speed. Many scholars have studied this problem, such as Kim et al. [33] modeling the discrete event behavior of tools through timing event diagrams. A feedback controller has been developed for single-arm and dual-arm cluster tools, which can meet the time limit by adjusting wafer delay. So that it does not exceed the prescribed limits. There are also studies on the acceleration of handling robots, such as Wang et al. [34], in order to solve the problems of large distance, high speed and high precision trajectory tracking. In order to ensure the stability of wafer transport robot, combined with the new section search method, a fifth-order polynomial interpolation method is proposed to plan the motion process of wafer transport robot and realize the optimal planning curve. And its effectiveness is proved, as shown in Figure 8B. Some scholars have studied the trajectory planning of flexible joint vibration suppression, which is also of great reference significance to the trajectory planning of wafer handling robots. For example, Kim et al. [35] proposed several new methods of robot vibration suppression input shaping technology. That is, the optimal S-curve trajectory, fixed zero-vibration shaper and dynamic zero-vibration shaper. These methods can suppress various vibration modes of flexible joint robot under computational torque control based on rigid model. Each planning method has its own characteristics and is suitable for different motion modes, as shown in Figure 8C. The following year, Kim et al. [36] proposed a new trajectory planning method for online optimal trajectory planning of flexible joints, which is optimal in dynamic approach time and maintains the integrity of path tracking in high frequency planning and control cycles. This method is based on the well-known path constrained time optimal motion. We show that this trajectory can be approximated quickly by trapezoidal velocity curve, so as to produce a trajectory close to the time optimization, and only four times of robot dynamics calculation are needed for each path segment. Finally, it is successfully applied to the elastic model. To sum up, trajectory planning of robot is essentially a constrained optimization problem, and there will be different research fields of trajectory planning for different constraints and optimization objectives. In the application of wafer handling robot, it is necessary to ensure efficiency and run smoothly, and it is also necessary to consider the problem of flexible joints. Many scholars have done research independently from the above perspective, but few studies have integrated the above research into wafer handling robot. 3.2. Compliance Control Techniques Through reading a large amount of data at home and abroad, we can see that many researchers start with control algorithms to improve the control accuracy of the system, and then reduce vibration and improve the running speed. At present, the main methods used are improving the traditional PID control algorithm and merging with other algorithms and various modern optimal control algorithms, such as neural network control, synovial control, fuzzy logic control, adaptive control, and so on. However, for wafer handling robot, the joints have highly nonlinear dynamics and strong connection, so it is difficult to achieve good dynamic and stable performance by traditional methods, so it is necessary to comprehensively consider the application of control algorithms to robots. This review introduces several typical control methods in wafer handling robots. 3.2.1. Improved PID Control Technology In practical application, traditional PID control is still the mainstream control. However, for flexible variable stiffness joint, it is nonlinear in normal working state, so the control effect realized by traditional PID control method is not ideal. It is necessary to add other control algorithms on the basis of traditional control, and then take the integrated algorithm as the algorithm of the whole system controller. Many scholars have done a lot of innovative research on the basis of classical PID control theory. In recent years, with the development of nonlinear control, decoupling servo control has attracted much attention. Many scholars have also applied it to wafer handling robots, which can obtain decoupling relationship through internal feedback loop and maintain system stability through external tracking error loop. Decoupling control is a very good choice for complex systems with various uncertainties and high frequency resonance, such as wafer handling robots. Han et al. [37] combined with the mechanical and software architecture of wafer handling robot, analyzed the robot kinematics by geometric method, decoupled the motion control from polar coordinates and joint space, and calibrated it by generalized least square inverse method. The center algorithm for motion control and wafer calibration of wafer handling robot in semiconductor manufacturing was studied, and the correctness, feasibility and effectiveness of the motion control and center correction algorithm of semiconductor robot were verified by experiments, which met the requirements of robot wafer transfer, as shown in Figure 9a,b. In addition, on the basis of PD control, He et al. [38] proposed a new coordinated control method based on decoupling servo control to design for wafer handling robot. As the basis of decoupling servo control, the dynamic model of wafer handling robot is obtained by Newton-Euler equation and Lagrange equation. These two methods confirm the validity of the kinetic equation. The inverse dynamics of the robot is analyzed by feedback linearization, and the experimental results show that the control method can effectively reduce the position tracking error. The performance meets the requirements of high speed and high precision in wafer processing, as shown in Figure 9c. Other researchers combine visual systems and adopt servo control strategies. Such as Xiao et al. [39] and designed a control system. The system takes PLC as the control core, is responsible for the communication between PC and robot, and executes logic control instructions. He et al. [40] can realize precise positioning and control decoupling of wafers through visual alignment. The reliability of visual servo control strategy is verified by simulation and physical prototype of wafer handling robot. The performance meets the requirements of wafer processing process. In addition to decoupling servo control, singular perturbation theory is also very effective in controlling flexible joints. When the stiffness of flexible joints with variable stiffness is large enough, the model of flexible joints with variable stiffness can be regarded as a system with fast and slow time domains. Singular perturbation control methods are generally divided into two kinds. One method is to select the motor tracking error as the fast variable, and the corresponding one is to select the connecting rod variable as the slow variable. Another method is to choose the joint elastic force as the fast variable, and at the same time, the connecting rod variable is selected as the slow variable; Singular perturbation method has the advantage of order reduction, and its algorithm is simple and clear, which can greatly reduce the calculation workload when considering elastic structures and obtain better control effect. The trajectory tracking problem of variable stiffness joints has been studied by many scholars based on singular perturbation theory. In the control system design, the motor tracking error is selected as the fast variable, and Javadi [41] and others model the motor dynamics as the time delay input of the antagonistic variable stiffness actuator. In addition, the nonlinear transformation of the antagonistic variable stiffness actuator is introduced, and then the super-twist sliding mode control is used to achieve position and stiffness tracking at the same time. Combining nonlinear transformation with super-twisting sliding mode control, the dynamic tracking of joint position and stiffness is realized at the same time. Prediction-based feedback, together with some disturbance observers, involves estimating external disturbance to compensate for input time delay, which improves the robustness of the system and achieves success in weakening the effect of external disturbance. The schematic diagram is shown in Figure 10. Another research on selecting joints as fast variables, such as Roozing et al. [42], designed the distribution mode of driving force for simulating variable stiffness joints driven by antagonistic tendons, which improved the energy efficiency. A new asymmetric antagonistic driving scheme is developed and controlled, which is characterized by large energy storage capacity and effective motion execution. Asymmetric design consists of two driving branches, and its power is transmitted to a single joint through two flexible elements with different stiffness and storage capacity characteristics. Given design parameters and control requirements, the selection criteria of stiffness of the two components are expounded, and the principle is shown in Figure 11. In view of the above research, the differential equation can be dynamically decomposed into quasi-steady system and boundary layer system by taking instantaneous joint stiffness as the system time scale and combining singular perturbation theory; For the boundary layer system, the active damping coefficient is set, and the velocity error between the load end and the input end is fed back to the system input to reduce the influence of vibration on the system. However, the application of singular perturbation control method also has some limitations. When the stiffness of joints is large enough or the model of flexible joints is reduced, the effect is better, and there is still room for change in the application of wafer handling robot. There are also many researches on control based on posture kinematics model and dynamics model. A promising vibration compensation method is to use feedforward control, which is realized by modifying the trajectory sent to the vibration machine. For example, Chou et al. [43] proposed a method to model the posture-dependent vibration of wafer handling robot as a function of its changing inertia and its experimentally determined joint stiffness and damping. The proposed modeling method first calculates the pose-dependent equivalent inertia of the system, then carries out some system identification experiments to determine the constant stiffness and damping coefficient, and then combines them to predict the dynamics under any pose. The attitude dynamics model of the wafer handling robot is used to design a feedforward tracking controller to compensate the attitude-dependent vibration of the robot. In order to enhance the tracking performance of the robot, it has a good effect, as shown in Figure 12A. For example, Yu et al. [44] proposed two control methods to improve the performance of wafer handling robot. Firstly, besides the basic three-loop PID feedback controller, torque offset is injected into the input of the robot, and feedforward is used to compensate the nonlinear dynamics of the robot. Secondly, a test-based optimization method is developed to adjust the gain of the feedback controller. The wafer vibration is measured and analyzed to score the system performance against the derived model. The related control effect is shown in Figure 12B. The comparison results indicate that position feedforward significantly enhances the robot’s tracking performance, while torque feedforward improves its anti-interference capabilities, leading to superior dynamic effects. The two proposed control schemes markedly enhance performance across various dimensions. To solve the problems of parameter uncertainty, unknown bounded friction torque, unknown bounded external disturbance and residual vibration, the ADRC method is firstly proposed by optimizing the classical feedback linearization control method. The core of this method is the extended state observer. The proposed technique does not need an accurate mathematical description of the system, because it is based on the on-line estimation and feedback of unmodeled dynamic elements and has good robustness [45]. For example, Guo et al. [46] combines linear extended state observer to control the position and stiffness of variable stiffness joints. A new conceptual model of antagonistic variable stiffness brake based on equivalent nonlinear torsion spring and friction damper is presented. A combination of linear extended state observer, sliding mode control and adaptive input saturation compensation law is used to design a robust tracking controller, and the position and stiffness of the torsion spring-based variable stiffness actuator against equivalent nonlinear are adjusted simultaneously. Under the proposed controller, the semi-globally uniform ultimate bounded stability of the closed-loop system is proved by Lyapunov stability analysis. Lukic et al. [47] proposed a cascade control structure for simulating the simultaneous control of position and stiffness of variable stiffness actuators driven by antagonistic tendons. The outer loop is composed of position and stiffness closed loops, and the inner loop is constructed by dominant dynamics, which can reduce system oscillation. Wisnu et al. [48] aims at the residual vibration caused by the high-speed operation of the transmission manipulator. They propose an integrated tool for parameter identification and vibration control of higher mode vibration systems. Using input shaping method to reduce residual vibration can help industrial personnel analyze and solve vibration problems. Experimental results show that the stability time and total motion time of the robot arm are significantly reduced. Kim et al. [49] analyzed from the wafer point of view and proposed an improved wafer alignment algorithm, which can obtain the relative distance between the wafer center and the robot hand, as shown in Figure 13. Some advantages of the proposed alignment algorithm include that it can be designed with low cost and less computational power. The proposed alignment algorithm is applied to the transfer robot of semiconductor processing, and the performance of the method is verified. The above research provides a basis for the application of semiconductor wafer handling robot, and provides a reference for more accurate control of wafer handling process. Many control methods are perfect enough, but due to the dynamics and nonlinearity of modeling, there are inevitably some inaccuracies in the model. After that, the research can continue to improve the model accuracy and nonlinear model parameter identification modeling direction. On the basis of accurate modeling, the application of the above control method will get twice the result with half the effort. 3.2.2. Intelligent Algorithm Control Because the wafer handling robot is composed of several joints coupled together, it is difficult to achieve good dynamic and stable performance by traditional methods. With the research of neural network, fuzzy control and artificial intelligence in the scientific community, new ideas are provided to solve many problems in flexible robot control. Intelligent control is mainly applied to relatively complex systems such as parameter uncertainty and structural uncertainty, as well as linear systems and deterministic systems with large time constants and large pure time delay. This is in line with the application of wafer handling robot, although these control methods are very complex, but there are still many scholars have done related research. By adjusting fuzzy parameters, the control based on fuzzy theory has the self-learning ability to discard the nonlinearity and uncertainty of the system, and can estimate the feedback linearized control. Hybrid fuzzy logic controller mainly depends on its control fuzzy rules and fuzzy membership function. Pizarro [50] proposed a strategy of fine-tuning membership function using genetic algorithm to control the tracking position of robot with two joints. The effectiveness of the control strategy is proved by comparing the simulated unoptimized controller with the optimized controller using genetic algorithm, as shown in Figure 14A. Fateh et al. [51] proposed a fine-tuning fuzzy control for robots. With the help of this controller, the trajectory tracking control simulation and experiment are carried out by using the articulated electric robot manipulator. Convergence analysis can track the target path effectively and robustly. Although this method is simple in calculation, its stability is guaranteed. The simulation results show that the fine-tuning fuzzy control is superior to the PD fuzzy control in the set-point and tracking control. However, unlike adaptive fuzzy control which adapts to all rules, fine-tuning fuzzy control adapts to only one fuzzy rule. Other adaptive control technologies can also control flexible joints, such as Trumic et al. [52]. For the problem of joint stiffness and position synchronization control based on variable stiffness joints, based on the existing adaptive control theory, an innovative control method composed of adaptive compensator and dynamic decoupler is introduced to develop a closed-loop control technology of joint stiffness. This method does not require higher position derivatives or any stiffness derivatives. The solution has been validated in multiple use cases, and comparisons reveal more powerful position and stiffness tracking skills. Adaptive control can also carry out joint control. For example, Liu et al. [53], This paper discusses a nonlinear direct joint control scheme, which is used to manipulate flexible payloads through single-arm systems with disturbances, parameter uncertainties and output constraints, and can avoid system overflow instability. Therefore, a new control method based on nonlinear disturbance observer is proposed to enhance its disturbance attenuation ability. In order to deal with the system parameter uncertainty, adaptive direct joint control is applied to approximate the unknown parameters, which does not need endpoint boundary control. Then, barrier Lyapunov function is used to eliminate the influence of yield restriction. Semigroup theory and invariance principle extended to infinite dimensional systems strictly prove the asymptotic stability of closed-loop systems and have good performance, as shown in Figure 14B. Neural network can solve the problem of unknown nonlinear function well. For the control problems caused by inaccurate modeling, external disturbance, joint flexibility and other uncertainties, You et al. [54] proposed a neural network control method based on H ∞. According to the singular perturbation theory, a robust controller based on neural network is used to decouple the dynamic model of rigid body. A feedback controller based on velocity difference is designed to compensate the angle error caused by joint flexibility. Experimental simulation results verify the effectiveness of the proposed control method, which has good engineering application value, as shown in Figure 14C. Rossomando Soria [55] proposed a neural sliding mode dynamic control in discrete time domain. The weight of radial basis function can be adjusted by online adaptation law. Sliding surface can limit the adaptive law and compensate by sliding surface to eliminate the approximate error introduced by neuron controller. There is similar adaptive neural backstepping integral sliding mode control based on disturbance observer, such as Phan et al. [56] integrating adjustable stiffness rotary actuator. The system is affected by unknown system dynamics, external disturbance and variable stiffness, so a disturbance observer is introduced to compensate for the influence of approximation errors caused by variable stiffness, disturbance and neural network. The effectiveness and feasibility of the proposed controller are proved by simulation and experiment. However, the neural sliding mode dynamic control theory has not been integrated, the parameter setting is difficult, and the application field is limited. Some scholars have integrated neural network, fuzzy control, adaptive control and PID control into one method. For example, Son et al. [57] proposed a new control system combining adaptive feedforward neural controller and PID controller, which uses actuators to control the joint angle position of robot. The inverse neural model is used to dynamically identify all the nonlinear features of the robot and the improved differential evolution algorithm is used to optimize it. Secondly, combined with an inverse neural model, the model provides applications from the desired joint position and conventional PID controllers to improve accuracy and reject steady-state errors in joint position control. Finally, a new adaptive back propagation algorithm for fuzzy systems is proposed to update the weights of the inverse neural model online, so as to better adapt to the interference and dynamic changes in its operation. Compared with the traditional control methods, the experimental results show that the proposed control scheme has better performance and advantages, but the control algorithm is very complex and there is room for simplified application. 3.3. Discission High safety, high stability and high reliability are three important factors for the development of high-performance robots in the future [58]. Under the background of high requirements of human-computer interaction, designing a better control method with variable stiffness and flexibility of robots can achieve the balance between speed, accuracy and vibration suppression during the movement of wafer transfer robots. The existing methods have their own advantages and disadvantages: Trajectory planning optimization: Considering the flexible transmission mode of wafer handling robot, it is necessary to combine the flexible dynamic model for trajectory planning. Fifth-order polynomial interpolation and S-curve programming significantly improve the smoothness of motion, but the residual vibration of flexible joints still needs to rely on feedback control compensation. The dynamic zero vibration shaper has remarkable effect at fixed frequency, but its adaptability to time-varying stiffness is insufficient, so it needs to be combined with on-line adjustment mechanism in practical application. Control algorithm: The existing flexible joints with variable stiffness of robots still have some problems in joint structure, variable stiffness range, dynamic stiffness identification and control methods. Decoupling servo control simplifies the control structure through inner loop decoupling, and has verified high tracking accuracy in robots, but there are still steady-state errors in systems with strong nonlinear coupling and sensitive time-varying parameters. Intelligent algorithms (such as neural network and fuzzy control) are outstanding in dealing with nonlinearity and uncertainty, but their computational complexity is high, which makes it difficult to meet the requirements of microsecond real-time control. For example, although the adaptive fuzzy-PID hybrid controller has improved robustness, the industrial deployment cost has increased sharply. Feedforward compensation and ADRC pre-compensate the vibration through the dynamic model, which significantly reduces the tracking error, but requires extremely high model accuracy and requires frequent calibration in practical applications. Industrial application feasibility: Existing intelligent algorithms mostly stay in the simulation stage, and industrial scenarios prefer feedforward-feedback composite control based on physical models. In the future, it is necessary to develop lightweight AI controllers supported by edge computing, and combine digital twins to realize online model updates to take into account accuracy and real-time performance, which is difficult to apply in practice.