Accurate 6-DoF pose estimation of spacecraft remains a critical challenge under extreme lighting conditions and strict computational constraints, and nonlinear geometric transformations in space applications. To address this, we propose the Spatial Stereo-Distributed Feature Network (SSFN), an end-to-end deep learning framework that combines orientation-aware soft classification with spatial feature fusion for robust and precise pose estimation using a lightweight monocular camera. The method operates without prior geometric models or stereo vision setups. At its core, a Keypoint Detection Network (KDN) based on a CSPNet backbone employs multi-task learning to predict 2D keypoints with associated uncertainty measures, reinforced by coplanarity constraints to improve geometric consistency. By incorporating virtual ray parameterization and depth-agnostic scaling, the framework achieves camera-parameterized 3D reconstruction and fuses geometric and appearance features, enabling effective handling of non-coplanar keypoints and significant depth variations without explicit depth input. For orientation estimation, a probabilistic soft classification module discretizes continuous orientations into Gaussian-distributed bins. This representation is aggregated via VLAD and refined through weighted leastsquares optimization, producing quaternion outputs with enhanced smoothness and robustness to noise. Extensive evaluations on SPEED, URSO, and LINEMOD datasets, supported by semi-physical simulations, demonstrate SSFN's state-of-the-art performance: it achieves mean translation and rotation errors of 0.0045 and 0.0096 on SPEED, a localization error of 0.47 and angular error of 5.6 degrees on the challenging URSO Soyuz_hard sequence, and 98.76 % accuracy under a 5 cm/5 degrees threshold on LINEMOD. These results confirm SSFN's efficacy in complex illumination and depth-varying scenarios, underscoring its potential as a reliable vision-based solution for autonomous spacecraft navigation.

This paper addresses the position and attitude control of combined spacecraft in on-orbit servicing missions, taking into account model parameter uncertainties, unknown external disturbances, and fuel-optimal constraints. A novel flexible prescribed-performance optimal backstepping controller without initial constraints is proposed by incorporating an Actor-Critic-Identify neural network architecture. First, a dynamic model of the combined spacecraft is established, with all uncertainties treated as lumped disturbances. To improve transient performance and remove initial value constraints, a flexible prescribed performance function is designed, which accommodates input saturation and decouples settling time from both initial states and controller parameters. Subsequently, a steady-state performance optimized Identify weight adaptation law is employed for rapid and accurate estimation of the nonlinear lumped disturbances. For fuel optimization, a simplified Actor-Critic adaptation law is developed, eliminating the need for complex step-by-step derivations while ensuring weight convergence. The uniform ultimate boundedness of the closed-loop system is proven using Lyapunov theory. Numerical simulations and semi-physical experiments verify the proposed method’s advantages in both steady-state and transient performance, as well as its applicability to on-orbit implementation.

Large flexible appendages can introduce persistent low-frequency vibrations that significantly degrade the attitude pointing accuracy of modern spacecraft. To address this issue, this paper proposes a hierarchical cooperative control framework for large flexible spacecraft that combines contraction-based tube model predictive attitude control with reinforcement learning-based active vibration suppression. First, a nominal model predictive controller is designed to generate reference attitude trajectories under input constraints. Then, an ancillary controller is developed based on contraction theory to drive the disturbed state toward the nominal trajectory. To improve real-time implementability, a Multidimensional Taylor Network (MTN) is employed to approximate the minimal geodesic required by the contraction-based control law, thereby reducing the online computational burden. In parallel, an Actor--Critic-based active vibration controller is introduced to attenuate flexible modal responses and reduce the adverse coupling effects of appendage vibrations on attitude motion. Lyapunov analysis establishes the robust convergence property of the attitude-control subsystem under bounded disturbances, while the cooperative effect of the vibration controller is validated through comparative simulations. Numerical results show that the proposed framework achieves faster attitude convergence and more effective vibration attenuation than the baseline methods under nominal and perturbed conditions.

In the field of spacecraft pose estimation, complex illumination variations and limited computational resources pose critical challenges. Supervised learning methods achieve higher accuracy but require massive labeled data for training, which is extremely costly to acquire for space scenarios. Meanwhile, traditional nonlearning methods, though label-free, suffer from significant performance degradation under complex illumination variations. This article proposes an end-to-end unsupervised framework for bidirectional optimization of pose estimation and 3D reconstruction. First, an unsupervised multiview stereo network is designed using a two-stage comparative learning strategy, which is capable of accurately reconstructing the target object from multiview images to obtain depth information. Then, a feature fusion method based on the attention mechanism is proposed to fuse the depth information into the pose estimation network so as to enhance the robustness of pose estimation. In addition, spacecraft pose constraints are obtained from the pose estimation, which in turn constrain the 3D reconstruction. Finally, the probabilistic inference capability of the Bayesian network is utilized to adapt to different lighting environments. The proposed method eliminates the need for large labeled datasets, suits resource-limited spacecraft, and enables efficient pose estimation using lightweight monocular cameras. Experiments on Spacecraft Pose Estimation Dataset, Unreal Rendering Spacecraft in Orbit, and a semiphysical platform show state-of-the-art (SOTA) accuracy (42.4% lower median rotation error than SOTA) and real-time capability (19 ms/frame).

This paper addresses the challenge of formation control and obstacle avoidance for spacecraft subject to unknown disturbances by proposing a flexible prescribed-time optimal formation control framework based on the Actor-Critic-Identifier network and attention mechanisms. First, a formation consensus error is constructed to guarantee configuration maintenance and velocity synchronization. Subsequently, a potential field function is developed to ensure obstacle avoidance among multiple spacecraft and between spacecraft and obstacles, while eliminating local minima. Concurrently, the prescribed-time performance function is flexibly adjusted according to the potential field, thereby mitigating conflicts between performance constraints and obstacle avoidance. Then, based on the optimal backstepping control theory, an Actor-Critic network is employed to approximate the calculation of the Hamilton-Jacobi-Bellman equation in the optimal problem. Furthermore, an attention-based Identifier network is proposed to estimate unknown external disturbances, which enhances both accuracy and estimation speed. Lyapunov stability theory is utilized to establish the stability analysis of the closed-loop formation system. The effectiveness and superiority of the proposed strategy are demonstrated through numerical simulations.

The complex surface configuration, tumbling motion, and surrounding debris environment of failed spacecraft pose significant risks and challenges to on-orbit capture and servicing missions. To ensure both safety and mission success under such conditions, this paper proposes a dual-layer model predictive control (MPC) strategy. In the upper layer, a Control Lyapunov Function (CLF) is employed to guarantee convergence, while a relaxation variable is introduced to prevent overly strict constraints, ensuring fast and flexible convergence of relative position and attitude. In the lower layer, Control Barrier Functions (CBFs) are used to impose strict safety constraints on multiple dynamic obstacles. A deadlock detection and adaptive reference adjustment mechanism is introduced between the two layers to effectively avoid local optima. Simulations under multiple initial conditions demonstrate that the proposed method achieves millimeter-level position accuracy and 0.03 degrees attitude synchronization, while effectively avoiding collisions with the target body, solar panels, and debris under external disturbances. The approach maintains near-global optimality while emphasizing safety, providing a feasible and robust solution for micro-nano satellite on-orbit servicing missions.

To address the insufficient predictive accuracy of traditional deployment torque models for inflatable tubes used in spacecraft deployable systems, this study proposes a refined analytical model that incorporates an incompleteellipse-parabola folding-contact surface. First, an incomplete elliptical cross-section hypothesis is introduced to describe the folded configuration of the inflatable tube, and a parabolic boundary is further adopted to represent the geometry of the contact region. This approach overcomes the limitation of conventional elliptical models, which fail to simultaneously account for the membrane stress-induced torque and the compressive torque from the contact interface. Based on the proposed geometric model, the stress distribution along the cross-sectional boundary is derived using membrane theory, from which the deployment torque is subsequently calculated. Finally, the analytical results are validated through static torque measurements and dynamic deployment experiments conducted on a planar air-bearing platform. The experimental results demonstrate that, within the angular range of 65 degrees-160 degrees, the proposed model achieves an average torque prediction error of only 5%, representing a 16% improvement over conventional models. Moreover, the predicted torque variation exhibits excellent agreement with the experimental trend. In addition, for a dumbbell-type inflatable dual-satellite system, the dynamic deployment behavior simulated using the proposed model closely matches the experimental observations, with an average deployment time error of less than 0.5s. These findings indicate that the proposed model can accurately simulate the deployment process of inflatable dual-satellite systems and provides a theoretical foundation for the development of future large-scale inflatable deployable space structures.

Compared with single-spacecraft capture schemes, multi-spacecraft on-orbit servicing offers significant advantages in decentralization, cost, and mission scalability. However, real-time trajectory planning for mutual collision avoidance, combined with the complex geometry of a failed spacecraft, imposes stringent requirements on the obstacle-avoidance and control algorithms of the servicers. To address these challenges, this study proposes a distributed cascaded model predictive control (MPC) architecture for a servicing satellite cluster. Each servicing spacecraft estimates the neighbors' motion trajectories at the current time step using historical information, thereby enabling inter-spacecraft collision avoidance within the swarm and realizing decentralized control; for an individual servicing spacecraft, in the upper layer, a control Lyapunov functions (CLFs)-based MPC guarantees closed-loop convergence while introducing relaxation variables to alleviate over-conservativeness, thereby enabling rapid and flexible convergence of the relative position and attitude. In the lower layer, control barrier functions (CBFs) are incorporated to rigorously enforce safety constraints against target-induced dynamic obstacles, and are further augmented by velocity-obstacle (VO) constraints and position-compatibility constraints for real-time online trajectory optimization, achieving inter-satellite deconfliction while reducing computational complexity in high-dimensional scenarios. A deadlock-detection and reference-reshaping mechanism is additionally embedded between the two layers to prevent entrapment in local minima. Simulation results demonstrate axial docking accuracy on the order of 0.1 mm and attitude synchronization accuracy of 0.05 degrees for the servicing cluster under complex constraints and cluttered obstacle environments. The proposed method ensures servicer safety and achieves near-globally optimal trajectories, providing a theoretical basis for multi-spacecraft attitude takeover, on-orbit refueling, and cooperative debris-removal missions. (c) 2026 COSPAR. Published by Elsevier B.V. All rights are reserved, including those for text and data mining, AI training, and similar technologies.

Star sensor is disturbed by strong straylight, which increases the gray level of the captured star map, and this leads to invalid detection of star points and affects the high-precision location of the centroid. To address this issue, we propose a star centroid localization method based on gradient-oriented multi-directional local contrast enhancement. First, the background gray level distribution patterns of star sensors under various actual straylight interference conditions are analyzed. Based on this analysis, a background imaging model for complex operational scenarios is established. Finally, simulations are conducted under complex conditions with straylight images to test the star point detection rate, false detection rate, centroid localization accuracy, and statistical significance testing. The results show that the proposed algorithm outperforms the TOP-HAT, MAX-BACKG (Max-Background Filtering), LCM (Local Contrast Measure), MPCM (Multiscale Patch-Based Contrast Measure), and CMLCM (Curvature-Based Multidirectional Local Contrast Method for Star Detection of Star Sensor) algorithms in terms of star point detection rate. Additionally, the RMSE centroid localization error is achieved with 0.1 pixels, demonstrating its ability to effectively locate star centroids under complex conditions and meet certain engineering application requirements.

Aiming at the problem that the precision of high-precision relative pose adjustable docking mechanisms is susceptible to collision forces, a model predictive-based compliant control method for docking mechanisms was designed. This method first decomposes the collision force into the forces acting on each telescopic rod of the docking mechanism through the Jacobian matrix, establishing a collision environment model for a single telescopic rod. Subsequently, based on the characteristics of environmental forces, a model predictive controller (MPC) was integrated into the traditional impedance controller for optimization, with real-time adjustment of input-output weights in the MPC according to the ideal compliance process. Numerical simulations demonstrated that the optimized controller reduced collision forces compared to traditional impedance control. Physical collision tests further revealed that the optimized docking controller reduced collision force peaks by 50%, exhibited faster response to collision forces, and enhanced adaptability to complex collision environments, achieving compliant docking and protection of the docking mechanism. Copyright (c) 2025 The Authors. This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0/)

This paper addresses the challenges of uncertain model parameters and unknown external disturbances in the control process of the combined spacecraft. A Steady-State Performance-Optimized Prescribed Performance Adaptive Controller (SPO-PPAC) is proposed by integrating the time-varying Barrier Lyapunov Function (BLF) with the Prescribed Performance Function (PPF) and the neural network weight adaptive law optimized for steady-state performance. The Radial Basis Function Neural Network (RBFNN) is employed to approximate the lumped disturbances caused by model uncertainties and unknown external disturbances. The adaptive law for neural network weights is enhanced to improve steady-state performance. The uniform ultimate boundedness of all state variables in the closedloop system is proven using Lyapunov stability theory. Numerical simulations demonstrate the advantages of the proposed algorithm in transient and steady-state performance. Semi-physical simulation experiments further illustrate the higher convergence precision of the (c) 2025 COSPAR. Published by Elsevier B.V. All rights are reserved, including those for text and data mining, AI training, and similar technologies.

The spacecraft microgravity simulation air-bearing platform is a crucial component of the spacecraft ground testing system. Special disturbances, such as the flatness and roughness of the contact surface between the air bearings and the granite platform, increasingly affect the control accuracy of the simulation experiment as the number of air bearings increases. To address this issue, this paper develops a novel compensation control system based on Active Disturbance Rejection Control (ADRC), which estimates and compensates for the disturbing forces and moments caused by the roughness and levelness of the contact surface, thereby improving the control precision of the spacecraft ground simulation system. A dynamic model of the multi-air-bearing platform under disturbance is established. A cascade ADRC algorithm based on the Linear Extended State Observer (LESO) is designed. The Gauss–Newton iteration method is used to identify the parameters of the sliding friction coefficient and the tilt angle of the air-bearing platform. A full-physics simulation experimental platform for spacecraft with rotor-based propulsion is constructed, and the proposed algorithm is validated. The experimental results show that on a marble surface with a flatness of grade 00, an overall tilt angle of 0–1 degrees, and a surface friction coefficient of 0–0.01, the position control accuracy for the simulated spacecraft can reach 1.5 cm, and the attitude control accuracy can reach 1°. Under ideal conditions, the identification accuracy for the contact surface friction coefficient is 2 × 10−4, and the recognition accuracy for the overall levelness of the marble surface can reach 1 × 10−3, laying the foundation for high-precision ground simulation experiments of spacecraft in multi-air-bearing scenarios.

Dynamically modeling the flexible characteristics of large-scale jointed composite spacecraft is challenging. In this study, a dynamic modeling method for rigid–flexible composite spacecraft is proposed based on the absolute nodal coordinate formulation (ANCF). First, the spacecraft in the jointed composite is simplified as a rigid body, and the docking mechanisms between spacecraft are approximated using the fully parameterized beam model. Next, regarding the constraints between the beam and the rigid body, the beam’s absolute nodal coordinates are converted into rigid body coordinates. This allows the dynamic equations to be simplified using independent coordinates, reducing the model dimension. Finally, system damping is increased through the mean stress noise reduction method, which suppresses high-frequency components in the dynamic model and further reduces the rigidity of the dynamic equations for the composite body. This modeling method decreases the complexity of the composite body dynamics and avoids the difficulty of solving algebraic–differential equations exhibited by Lagrange multiplier methods, facilitating numerical simulations. The proposed method is applicable to both tree and mesh topologies. MATLAB simulations demonstrate that the proposed dynamic model alleviates the dimensionality disaster caused by conventional algorithms, significantly reducing computation time. The simulation results are consistent with ADAMS. The proposed model exhibits displacement errors less than 1 mm, highlighting its efficiency and accuracy.

To address the limitations of current micro-nano satellites in on-orbit monitoring of non-cooperative targetssuch as inadequate adaptability in attitude control dynamics, lack of coupled orbit attitude control strategies, and insufficient handling of state constraintsthis paper establishes a six-degree-of-freedom relative motion dynamics model for spacecraft approaching non-cooperative targets. The model systematically incorporates practical engineering constraints, including input saturation and velocity limits. By integrating Control Barrier Functions (CBF) to formulate obstacle avoidance strategies, a robust integrated orbitattitude control method based on Model Predictive Control and Control Barrier Functions (MPC-CBF) is proposed. The proposed method provides both theoretical support and a technical framework for on-orbit servicing missions involving micro-nano satellites, demonstrating strong potential for practical applications.

This paper investigates the problem of angular momentum control and planning for control moment gyroscope(CMG) arrays in rigid spacecraft attitude control systems using deep reinforcement learning (DRL). Specifically, a DRL-based angular momentum control strategy is proposed for spacecraft attitude control systems employing multiple CMGs as actuators. The twin-delayed deep deterministic policy gradient (TD3) algorithm is used to perform online learning and policy updates based on environmental feedback. This approach eliminates the need for precise mathematical models and iterative parameter tuning. This enables the CMG system to perform angular momentum planning and facilitates rapid and high-precision spacecraft attitude maneuvers and control through angular momentum exchange. Simulations were performed to analyze spacecraft attitude maneuvers and stabilization under various scenarios, focusing on the angular momentum control process of a pyramidal single gimbal CMG (SGCMG) array. The results demonstrate that the proposed method effectively achieves large-angle attitude maneuvers and stable attitude maintenance, both in ideal conditions and in the presence of nonlinear disturbances. During large-angle maneuvers, the spacecraft’s attitude estimation using MRPs converges in less than 1 min, and the convergence accuracy during attitude-holding reaches the order of 10−3. Moreover, the approach fully leverages the output characteristics of the CMG system and achieves robust performance and accuracy, even under conditions with significant noise and disturbances.

In order to combine micro-nano satellites into ultra-large variable structure spacecraft, the design of relative attitude repeatedly adjustable docking mechanism with high precision is the key problem. Based on Stewart mechanism, a kind of high-precision relative attitude repeatedly adjustable docking mechanism was designed. Aiming at the problem of relative attitude high-precision adjustment after docking locking, the transfer mechanism of attitude adjustment error of docking mechanism was analyzed, and the transfer model of attitude adjustment error was established. The theoretical attitude adjustment accuracy was obtained through simulation calculation. By forming a closed-loop control system with the laser sensor and Stewart mechanism, the error compensation algorithm of the pose control system was designed, the high-precision adjustment of the relative pose was realized. The dynamic model was established in the scenario of two-module relative attitude adjustment, and the driving force variation law of each telescopic rod motor was analyzed according to the motion trajectory of attitude adjustment. Simulation analysis and real object attitude adjustment test show that the weight of the docking mechanism is only 2.43kg,the radius of the envelope is only 10cm,the accuracy of the three-axis attitude adjustment can reach +/- 0.02 degrees,and the maximum attitude adjustment load of the docking mechanism is 65N,which can be applied to the precise allostery of micro-nano assembly spacecraft.

To address the limitations of inadequate real-time performance and robustness encountered in estimating the pose of non-cooperative spacecraft during on-orbit missions, a novel method of feature point distribution selection learning is proposed. This approach utilizes a non-coplanar key point selection network with uncertainty prediction, pioneering in its capability to accurately estimate the pose of non-cooperative spacecraft, thereby representing a significant advancement in the field. Initially, the feasibility of designing a non-coplanar key point selection network was analyzed based on the influence of sensor layout on the pose measurement. Subsequently, the key point selection network was designed and trained, leveraging images extracted from the spacecraft detection network. The network detected 11 pre-selected key points with distinctive features and was able to accurately predict their uncertainties and relative positional relationships. Upon selection of the key points exhibiting low uncertainty and non-coplanar relative positions, we utilized the EPnP algorithm to achieve accurate pose estimation of the target spacecraft. Our experimental evaluation on the SPEED dataset, which comes from the International Satellite Attitude Estimation Competition, validates the effectiveness of our key point selection network, significantly enhancing estimation accuracy and timeliness compared to other monocular spacecraft attitude estimation methods. This advancement provides robust technological support for spacecraft guidance, control, and proximity operations in orbital service missions.

The utilization of monocular vision for non-cooperative spacecraft pose estimation has been significantly researched in space target monitoring, on-orbit servicing, and satellite maintenance. The challenge lies in addressing the cross-domain variations in shape, texture, lighting, and motion patterns between simulated and real captured images. To tackle this issue, a novel domain adaptation 6DoF pose estimation algorithm is proposed to extract the geometric and semantic consistency between cross-domain training and testing datasets. Experimental results demonstrate that our pose estimation method achieves state-of-the-art performance on the SPARK2024 dataset.

In large-scale satellite constellations, the efficiency of inter-satellite communication is paramount. Traditional topology control strategies, such as the Manhattan configuration, provide stable links but can result in indirect communication paths, affecting the efficiency of information transfer. This paper addresses this issue by proposing an innovative “3 + 1” dynamic topology control scheme. The scheme retains three static links determined by the relative angular velocity and acceleration while introducing a dynamic link based on distance and angular velocity constraints to optimize the link duration and overall network communication efficiency. To address the complexity of matching dynamic links, this paper introduces an elite strategy-based maximum weighted matching algorithm for general graphs. Compared to traditional greedy algorithms, our proposed algorithm significantly improves the link duration and topological stability. Through simulation experiments comparing communication delays between Xiamen and Los Angeles, our results show that the proposed dynamic link scheme substantially reduces the average delay, enhancing the efficiency and flexibility of inter-satellite communication. This research not only extends the duration of inter-satellite links but also provides new perspectives and methodologies for further studies on inter-satellite topology control strategies.