Inertia parameter identification via neural network and reinforcement learning for combined spacecraft attitude takeover control

The attitude takeover control of the combined spacecraft after capturing the non-cooperative target has become one of the key technologies for on-orbit servicing and maintenance of spacecraft. However, a complete takeover control process necessitates first identifying the inertial parameters of the combined spacecraft and then conducting the attitude takeover control. Yet, the harsh and complex space environment, compounded by significant angular velocity measurement noise, makes such inertia parameter identification extremely difficult, posing formidable challenges to this task. In order to accurately identify the inertia parameter, we propose an LSTM-based neural network inertia parameter identification approach. This approach can complete the identification within several milliseconds while maintaining a lightweight architecture, thereby rendering it more suitable for on-orbit scenarios with constrained resources. A reinforcement learning-based control strategy is proposed for the attitude takeover control of the combined spacecraft. This strategy enables real-time and accurate identification of the inertia parameter and features thoughtfully designed state representations and reward functions that collectively enhance robustness. Extensive experiments are conducted under diverse measurement noise conditions to simulate different on-orbit mission scenarios, and the findings consistently underscore the proposed framework’s performance advantages over existing identification approaches. The RL-based controller demonstrates strong robustness to variations in the inertia parameter, consistently maintaining high-precision attitude control and rapid, stable recovery across different configurations. The simulation results collectively validate the efficacy of the proposed integrated framework. Specifically, the neural network exhibits high accuracy in estimating the combined spacecraft’s inertial parameters, while the RL-based controller achieves precise attitude regulation with minimal overshoot and fast settling times. © 2026 Elsevier Masson SAS.