State estimation and dynamic compensation cooperative sliding mode control for space flexible manipulators Authors: Juan Wang, Aiping Wang, Beibei Wang, Yuanqi Chen, Liping Li Published: 2026 ABSTRACT AbstractIn response to the issue of total interference suppression in space flexible manipulator systems, a state estimation and dynamic compensation cooperative sliding mode control strategy is proposed, based on a neural network-based disturbance observer and a sliding mode observer. This approach leverages the approximation capability of neural networks to estimate and compensate for the system’s composite uncertainties in real-time. By combining a nonsingular terminal sliding mode control law, it ensures that the system’s state converges to the equilibrium point within a finite time, significantly improving the system’s tracking accuracy and robustness against disturbances. Considering the practical scenario where the system’s states may not be fully measurable, a sliding mode observer is designed to accurately reconstruct the system’s unknown states. Based on these estimated values, a cooperative sliding mode controller is designed. This cooperative control scheme effectively addresses the control challenges when the system’s states are partially unmeasurable. Simulation results show that the system maintains excellent dynamic performance and stability, even in the presence of nonlinear factors such as friction.Keywords: Neural network disturbance observer, Sliding mode observer, Nonsingular terminal sliding mode control, Cooperative controlSubject terms: Engineering, Mathematics and computing Introduction With the rapid development of aerospace technology, space missions such as on-orbit servicing and space debris removal have imposed extremely high demands on the performance of robotic manipulators1–3. Due to the use of lightweight materials, space manipulators exhibit flexible vibration characteristics, which significantly increase the difficulty of achieving high-precision control4. To address these challenges, cooperative control strategies in the field of unmanned systems have provided new insights in recent years: Yang et al.5 proposed a pseudo-trajectory homotopy method for information collection by underwater unmanned marine vehicles; Jin et al.6 investigated path planning for collaborative data acquisition by heterogeneous unmanned aerial vehicles; Zhu et al.7 designed a hybrid event-triggered control strategy for cooperative positioning of multiple unmanned surface vehicles; Li et al.8 studied prescribed-time consensus for multi-agent systems under input saturation; and Huang et al.9 proposed a formation safety control approach combining graph theory with reinforcement learning. However, beyond cooperative issues, the control performance of space flexible manipulators is further constrained by compound uncertainties such as unmodeled dynamics, parameter perturbations, and time-varying disturbances. Additionally, nonlinear factors like joint friction and backlash further exacerbate the difficulty of achieving high-precision control10,11. To cope with the aforementioned challenges, scholars domestically and internationally have proposed various control strategies. Traditional methods, such as PID control12, feature a simple structure but struggle to handle nonlinearities and time-varying disturbances. Fuzzy control13, while adept at processing imprecise information, relies heavily on expert knowledge for its performance. Conventional sliding mode control (SMC)14 is renowned for its strong robustness but suffers from inherent chattering. Adaptive control15 can adjust parameters online but often responds insufficiently to sudden disturbances. These methods find it difficult to simultaneously and effectively address parameter perturbations and complex external disturbances. If such uncertainties are not properly resolved, they will directly lead to a decline in system tracking accuracy and even threaten closed-loop stability. Recently, observer-based compensation methods have shown great promise. In terms of disturbance estimation, neural network-based disturbance observers16–18 have garnered significant attention for their strong nonlinear approximation abilities. Meanwhile, fuzzy disturbance observers19 and predefined-time20 or high-order nonlinear disturbance observers21,22 have also been developed to achieve rapid and precise disturbance estimation. Regarding state reconstruction, sliding mode observers23 and their high-order variants24,25 have been successfully employed to reconstruct unknown system states, leveraging their finite-time convergence characteristics. The integration of sliding mode differentiators with state observers has further improved the capability to extract derivative signals from noisy data26. At the controller design level, nonsingular terminal sliding mode control (NTSMC) addresses the singularity problem of traditional terminal sliding mode by introducing nonlinear sliding surfaces, ensuring finite-time convergence of the system, with continuous innovations in related designs27–29. Moreover, fast terminal sliding mode surfaces constructed using hyperbolic tangent functions can significantly accelerate convergence speeds when the system is far from the equilibrium point30,31. Although the aforementioned observers and advanced control methods have achieved success individually, organically integrating them to leverage their synergistic advantages remains a significant challenge. Existing combined schemes often suffer from structural complexity and heavy computational burdens. Disturbance observers and state observers are typically designed independently, failing to fully exploit their synergistic estimation advantages. Consequently, implementation costs and real-time performance in practical engineering applications have become bottlenecks that urgently need to be resolved. In light of this, this paper proposes a sliding mode control strategy featuring synergistic state estimation and dynamic compensation to overcome the aforementioned limitations. The core contributions of this work are as follows: We design an estimation framework where an RBF Neural Network (RBFNN) disturbance observer works in synergy with a Sliding Mode Observer (SMO). This architecture leverages the feedforward compensation capability of the RBFNN and the feedback estimation ability of the SMO to achieve precise estimation of compound uncertainties and finite-time reconstruction of unknown system states. We design an estimation framework where an RBF Neural Network (RBFNN) disturbance observer works in synergy with a Sliding Mode Observer (SMO). This architecture leverages the feedforward compensation capability of the RBFNN and the feedback estimation ability of the SMO to achieve precise estimation of compound uncertainties and finite-time reconstruction of unknown system states. We combined the aforementioned collaborative estimator with the nonsingular terminal sliding mode control (NTSMC) to ensure that the system state could converge to the equilibrium point within a finite time, even in the presence of nonlinear factors such as friction. We combined the aforementioned collaborative estimator with the nonsingular terminal sliding mode control (NTSMC) to ensure that the system state could converge to the equilibrium point within a finite time, even in the presence of nonlinear factors such as friction. Through comparative simulations with two controllers (denoted as RN and RSS), we comprehensively validate the superior performance of the proposed RSN scheme in terms of tracking accuracy, convergence speed, disturbance rejection capability, and control smoothness. Through comparative simulations with two controllers (denoted as RN and RSS), we comprehensively validate the superior performance of the proposed RSN scheme in terms of tracking accuracy, convergence speed, disturbance rejection capability, and control smoothness. Dynamic modeling of robotic arms Suppose the manipulator has only two flexible links, and the bending of each link occurs within a two-dimensional plane, as shown in Fig. 1. The lengths of the two links are denoted as and , and their corresponding masses are and . The position coordinates and represent the positions of the rigid body parts on each degree of freedom. The manipulator’s joints are connected through flexible elements, and the deformation function describes the bending and deformation of the flexible link at various positions along the link. Structure of the two-link flexible manipulator. For the flexible manipulator, assume that its deformation can be represented by a set of modal functions. The lateral displacement is represented as the sum of , and the deformation function can be expressed as: where is the generalized coordinate of the i-th mode, representing the amplitude of that mode, is the spatial distribution function of the i-th mode, and represents the inherent modes of the structure. The Lagrange equation for each generalized coordinate can be written as: where L is the Lagrangian, is the generalized coordinate, is the generalized velocity, is the generalized force, i.e., the torque applied at the joints, and t is time. The dynamic model of the manipulator control system is as follows: where is the vector of link displacements, is the vector of link velocities, is the vector of link accelerations, is the mass matrix, is the Coriolis matrix representing the inertial and velocity-dependent terms, is the gravitational term, is the motor torque, and is the velocity-dependent nonlinear frictional force. Equation (3) can also be expressed in state-space form as: Where , and . Main result Disturbance observer design The Radial Basis Function Neural Network Disturbance Observer (RBFDOB) is used to estimate and compensate for external disturbances and unmodeled dynamics. Given the significant nonlinear characteristics of the flexible manipulator, the RBFDOB can effectively estimate nonlinear disturbances, thereby enhancing the system’s robustness. As shown in Fig. 2, the RBFDOB is combined with sliding mode control to estimate the unknown disturbances in the system, further suppressing disturbances and improving control accuracy. Closed-loop posture control system structure with the RBF disturbance observer. A neural network observer is designed to approximate the real disturbance, such that: . It is assumed that the disturbance d can be represented as: Where , W is the ideal neural network weight matrix, is the basis function, and is the approximation error. The disturbance is estimated by the neural network as: where is the estimated weight. The observer dynamics are designed by defining the estimation error as , where the nominal acceleration is . The deviation between the actual acceleration and the nominal acceleration is caused by the disturbance: Thus, the disturbance estimation error is: The neural network weight update rule is: where is the neural network weight update gain matrix, which can be represented as , being the learning rate for the ith weight component, and is the radial basis function, typically modeled as a Gaussian function: where is the center of the ith basis function, and is the width parameter. To allow the RBF network to better adapt to time-varying disturbances, the centers can be adjusted online using the gradient descent method. The objective function is minimized to reduce the disturbance estimation error: The gradient with respect to is computed as: Using an online update rate gradient descent method, we obtain: where is the learning rate. Since d is unknown in practice, we can replace the error with the observable estimation error . The stability analysis of the neural network interference observer is conducted, and the Lyapunov function is defined as: By treating the first term as the kinetic energy of the observation error and the second term as a penalty on the observation weight error, the time derivative of V is obtained as: Substituting the error dynamics and the weight update law into the above gives: Using the properties of the system inertia matrix, one knows that is skew-symmetric, hence ; from the weight-update law it follows that . Substituting these into (16) yields: Since , and , we simplify to obtain: For the approximation error term : since C is skew-symmetric, the Coriolis term satisfies. Therefore: It follows that when , may be positive and the system can be temporarily unstable. However, by designing an appropriate robust term, the influence of can be suppressed, ensuring uniform ultimate boundedness. Sliding mode observer design The interference estimation provided by the nonlinear interference observer can help correct the errors caused by unmodeled dynamics or external disturbances, and improve the accuracy of state estimation. At this point, the accurate state estimation of the sliding mode observer (SMO) also helps the nonlinear disturbance observer to identify and compensate for disturbances more precisely. Now consider the friction term in the state-space representation of (4), defined as: where , and and denote the Coulomb and viscous friction at the ith joint, respectively. Assume this is a special case of a nonlinear system.For given and . For the state-space model in (4), the time-varying sliding-mode observer system32 takes the following form: where is the estimate of , and and are positive-definite constant diagonal matrices, i.e., , and are state-dependent matrices to be defined later. The expression of is defined as: The observation error e is defined as the difference between the process output and its estimate: The state error of the nonlinear observer is defined as . Introduce the state error,They are respectively expressed as ,, Subtracting the observer system from the nonlinear system yields the error system : Where is defined as: In the observer system Eq. (21), to emphasize using instead of , and since the state can be explicitly measured from the output signal y, we replace with y. Consequently, the gravity term is no longer a function of , because the term can be computed directly. The last term in (25) can be directly replaced by , where . This yields the following expression for: Introduce the function . When , the plane is associated with gravity. Its initial condition satisfies: so that the trajectory enters the plane . The behavior inside this reduced-order cluster can be formally derived using Filippov’s concept. The resulting dynamics are given by (24) and (25): From the sliding condition, the definition of is obtained as: Let , then: Due to the skew-symmetry of , the last term on the right-hand side of this expression is zero. Then, by arbitrary choosing and , the remaining terms can be made negative, verifying the condition: The matrices Q and can be chosen as diagonal positive-definite matrices, . Then calculate as: If the observer gain is large enough, the observation error converges to a bounded neighborhood B of the origin within a finite time, and the radius r can be arbitrarily reduced by increasing the minimum eigenvalue of . Theorem 1 Assume are as specified and the observer gain is sufficiently large. Then, the squared norm of the velocity error converges exponentially to the open set , defined as: where, the radius defines the accuracy of the observer and can be made arbitrarily small by increasing the smallest eigenvalue of the design matrix Q. Proof By deriving conditions on the matrix Q, and consequently on and , we ensure that P is positive definite. If all the leading sub-terms of matrix P are strictly greater than zero, then, due to the symmetry of P, all its eigenvalues are nonzero. When the matrix takes the form ,the expression of is: It follows directly that a necessary and sufficient condition for P to be positive definite is:. Now, by proving the existence of a positively bounded , we verify the statement given in Theorem 2. As long as is negative, will decrease. That is, for all , the following inequality needs to be verified: As long as , the radius is positive. The dependence of on the value of p can be obtained from the characteristic polynomial of P: Two sets of n equal eigenvalues are given. It can be easily proved by the solution of the second-order equation that as p increases, tends to 1. Therefore, the radius tends to zero as p increases. The proof is complete. Cooperative sliding-mode controller with state estimation and dynamic compensation: design and stability analysis Using the SMO, the state is predicted and a sliding term is employed to compensate model mismatch and disturbances, yielding the estimated state . The SMO’s estimated state is then used as the input to compute the neural-network–based disturbance estimate, producing the NN disturbance–observer estimation error . The principle block diagram of this composite control scheme for the flexible manipulator is shown in Fig. 3: Structure of the cooperative sliding-mode controller with state estimation and dynamic compensation. The overall scheme forms a closed-loop controller that leverages the strong robustness of sliding-mode control and the approximation capability of neural networks, ensuring accurate trajectory tracking in the presence of uncertainties and disturbances. Let the desired trajectory be . Define the tracking error ; based on the system input in (21), construct the sliding surface given in33 as follows: Where and . Taking the derivative of the sliding surface yields: Assuming ), then: For the space flexible manipulator, substitute the SMO-estimated values and the state error into the system, and design a finite-time sliding-mode control law that combines equivalent control with a switching control term based on the RBF neural-network disturbance observer and the sliding-mode observer termed RBFDOB-SMO-NTSMC (abbreviated RSN): To analyze the stability of the composite controller, consider both the observer error and the tracking error, and define the composite Lyapunov function: where and P is a symmetric positive-definite matrix. Differentiating and substituting the control law gives: where and . Choose ; then: According to the theory of finite-time stability, the system state converges to in finite time, after which the error e asymptotically converges to zero along the sliding surface. Simulation results To verify the feasibility of the proposed hybrid controller, a two-joint flexible manipulator is used as an example. MATLAB 2023b Simulink serves as the simulation platform. The desired trajectories of the two joints are set as and , respectively, and the simulation time is 15 s. The model parameters are designed as follows: where is the moment of inertia about the center of mass, is the mass of the ith link, is the distance to the center of mass of the ith link, is the length of the ith link, is the angle of the ith joint, and is the gravitational acceleration. It is assumed that , , and . The parameters of its robotic arm dynamic system are shown in Table 1 as follows: Table 1Parameters of manipulator.SymbolParameterSymbolParameterSymbolParameterSymbolParameter 1.5 0.8 1 0.8 1 1 1 1 g 9.8 0.8 1 Parameters of manipulator. The RSN control law proposed in this paper is denoted Controller 1, and it is compared against the following two controllers: Controller 2: To assess the role of the SMO in the proposed controller’s performance, the SMO module is removed, yielding a finite-time sliding-mode controller based on the neural-network disturbance observer (RBFDOB-NTSMC, abbreviated RN). The controller is given by: where the parameters are chosen to be the same as in Controller 1. Controller 3: To evaluate the accuracy contribution of the nonsingular terminal sliding surface in the proposed controller, a sliding-mode control with an exponential reaching law is adopted, resulting in a sliding-mode controller based on the neural-network disturbance observer and the sliding-mode observer (RBFDOB-SMO-SMC, abbreviated RSS). The controller is given by: where the parameters are set to . The parameter values of the controllers for the above three control schemes are shown in Table 2. Table 2Design parameters of each controller.Control schemeValueRSN RN RSS Design parameters of each controller. To effectively test the robustness and vibration suppression capability of the sliding mode observer in the presence of measurement noise, Gaussian white noise was added after the system ran for 10 s. The interference bandwidth was 10 Hz, the amplitude standard deviation was 0.55, and the amplitude approximate range was ± 1.65. To demonstrate the superiority of the sliding-mode observer, Fig. 4 compares the RN and RSN schemes. Comparison of joint responses between RN and RSN controllers. Figure 5 illustrates the trajectory tracking response curves for Joint 1 and Joint 2 under the RSS and RSN schemes. Combined with the settling time data presented in Table 3, it is evident that the proposed RSN scheme demonstrates a significant advantage in convergence speed. Comparison of joint responses between RSS and RSN controllers. Table 3Data comparison of controllers.Controller 1 (RN)Controller 2 (RSS)Controller 3 (RSN)Adjustment time for joint 1 angular displacement (s)0.800.900.75Adjustment time for joint 2 angular displacement (s)0.750.850.70 Data comparison of controllers. Figures 4 and 5 respectively show the trajectory tracking response curves of RN versus RSN, and RSS versus RSN for Joint 1 and Joint 2. To quantitatively evaluate the convergence speed of each controller, the settling time is defined as the shortest time required for the system response to first enter and remain within the ± 5% error band of the steady-state value. From Table 3, it can be seen that the settling times of the RSN scheme for Joint 1 and Joint 2 are 0.75 s and 0.70 s, respectively. Compared with the RN scheme (0.80 s and 0.75 s), these are reduced by 6.3% and 6.7%, respectively. Compared with the RSS scheme (0.90 s and 0.85 s), they are reduced by 16.7% and 17.6%, respectively. These results validate the advantage of the non-singular terminal sliding mode surface design in accelerating the system’s convergence speed: Although the RN scheme incorporates neural network disturbance compensation, it lacks the state reconstruction capability of a sliding mode observer, resulting in a slightly slower convergence speed; the RSS scheme employs an exponential reaching law, leading to a relatively smoother convergence process; in contrast, the RSN scheme, by utilizing the sliding mode observer to reconstruct system states in real time, combined with the finite-time convergence characteristic of the non-singular terminal sliding mode surface, achieves the fastest dynamic response. Figure 6 compares the acceleration tracking error curves of the three controllers for each joint. The magnitude and fluctuation degree of the acceleration error directly reflect the system’s vibration suppression capability. To further quantify this performance, Table 4 presents the vibration reduction effect percentages of each control algorithm on Joint 1 and Joint 2, defined as the attenuation rate of the vibration amplitude compared to the uncontrolled state. Trajectory error comparison among the three controllers. Table 4Comparison of damping effect.Control algorithmJoint 1 (%)Joint 2 (%)RBFDOB-NTSMC57.261.4RBFDOB-SMO-SMC93.391.7RBFDOB-SMO-NTSMC98.398.3 Comparison of damping effect. From the data in Table 4, it can be observed that the vibration reduction rates of the RN scheme for Joint 1 and Joint 2 are 57.2% and 61.4%, respectively. This indicates its limited capability in suppressing high-frequency vibrations induced by external noise and unmodeled dynamics. The RSS scheme demonstrates significant improvement, with vibration reduction rates reaching 93.3% and 91.7% for the two joints. This suggests that the introduction of the sliding mode observer greatly enhances the accuracy of system state reconstruction, thereby effectively suppressing vibrations caused by state estimation errors. The RSN scheme performs the best, achieving vibration reduction rates as high as 98.3% for both joints. Compared with the RN scheme, the vibration reduction rate is improved by an average of approximately 39%; compared with the RSS scheme, it is improved by an average of about 6%. This advantage is attributed to the synergistic mechanism of the RSN scheme: the sliding mode observer provides accurate state estimation, the neural network disturbance observer compensates for composite uncertainties in real time, and the non-singular terminal sliding mode surface ensures the smooth convergence of system states to the equilibrium point within a finite time. As shown in Fig. 7, the RSS controller suffers from high-frequency chattering due to the exponential reaching law, leading to pronounced torque oscillations. The RN controller reduces chattering slightly but exhibits larger torque amplitudes due to weaker disturbance rejection. In contrast, the RSN controller, benefiting from accurate disturbance estimation via the sliding-mode observer and the finite-time convergence property of the nonsingular terminal sliding surface, achieves smooth and properly scaled torque outputs. Control torque outputs of the three controllers. To sum up, the RSN scheme integrates RBFDOB-based model uncertainty compensation, SMO-based external disturbance estimation, and NTSMC-based finite-time convergence to form a multilayer disturbance-rejection architecture. It outperforms the comparative methods in tracking accuracy, response speed, robustness, and control smoothness. Particularly in terms of vibration reduction, the RSN scheme achieves an attenuation rate of nearly 98%, demonstrating its effectiveness in handling the complex vibration environment faced by space flexible manipulators during on-orbit servicing, and highlighting its significant potential for engineering applications. The RN controller, relying on a single observer, lacks robustness against compound disturbances, while the RSS controller cannot ensure finite-time stability due to its reaching-law design. The simulation results verify that the proposed “observer + finite-time control” strategy is effective and provides a promising approach for high-precision motion control. Conclusion In conclusion, this study verified the effectiveness of the collaborative strategy of state estimation and dynamic compensation in addressing the challenges of trajectory tracking control for robotic arms, providing a promising solution for high-precision and high-robustness motion control. Future work will proceed in the following directions. First, further optimizing the parameter adaptation laws for both the neural network and the sliding mode observer to reduce sensitivity to controller parameters. Second, investigating the application of this cooperative control framework to more complex scenarios, such as those involving input saturation, actuator faults, or on-orbit servicing tasks. Third, exploring online tuning of the composite controller using intelligent methods like reinforcement learning to further enhance the system’s adaptability and overall performance. Author contributions Aiping Wang (A.W.): Proposed the core research idea, designed the cooperative sliding mode control strategy, performed the theoretical analysis, drafted Figs. 1, 2 and 3, and wrote the initial version of the manuscript.Juan Wang (J.W.): Supervised the overall research direction and played a decisive role in the later stages of the study. Specifically, J.W. led the comprehensive data analysis, orchestrated the reconstruction of the paper’s structure, secured the funding support, and performed the critical revision and finalization of the manuscript.Beibei Wang (B.W.): Implemented the numerical simulations, processed the experimental data, and generated the remaining figures and results.Yuanqi Chen (Y.C.): Developed the dynamic model of the space flexible manipulator, provided essential technical resources, and assisted in the simulation setup.Liping Li (L.L.): Assisted in the stability analysis, validated the simulation results, and contributed to the polishing and editing of the manuscript. Funding This research was funded by the Innovation Capability Support Plan Project of Shaanxi Province (No. 2024ZC-KJXX-020), the Shaanxi Association for Science and Technology Youth Talent Support Program (No. 20230520). Data availability All data, models, and code generated or used during the study appear in the submitted article. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement All data, models, and code generated or used during the study appear in the submitted article.