TY - JOUR
T1 - Efficient Reinforcement Learning for Satellite Evasion Tasks in Limited Observable and Orbit-Constrained Environments
AU - Wei, Linsen
AU - Ning, Xin
AU - Lian, Xiaobin
AU - Yang, Kaiyuan
AU - Ni, Tao
AU - Chen, Feng
AU - Tang, Rugang
AU - Chen, Zhansheng
AU - Bo, Liteng
N1 - Publisher Copyright:
© 2026 Linsen Wei et al.
PY - 2026
Y1 - 2026
N2 - Reinforcement learning (RL) holds promise for the automated control of satellite orbits and the avoidance of spatial obstacles due to its ability to learn and adapt to dynamic environments through interaction, enabling autonomous, real-time decision-making for orbit adjustments. Satellite orbit control tasks manifest strong nonlinear characteristics, for which algorithms grounded on neural networks are more adept at handling. However, RL methods based on neural networks, which are highly nonconvex and nonsmooth functions, sometimes yield extreme action values that are not conducive to the orbit-constrained task at hand. In addition, in practical satellite task environments, limited observability is often faced. This limited observability not only increases the complexity of the algorithm but also affects the efficiency of identifying appropriate actions. To address this issue, we introduce a hierarchical parameterization mechanism to accelerate the search for appropriate actions. Hence, we introduce the integration of selfattention mechanisms with the Kalman filter into RL frameworks to accommodate scenarios with limited observability. Experimental results demonstrate a marked increase in the success rate of automated evasion maneuvers in environments with limited observability and orbital constraints, when facing pursuing satellites. Then, we performed a comparative analysis, applying Lambert guidance and RL-based guidance as control strategies for powerful confinement environment to validate our algorithm’s robustness.
AB - Reinforcement learning (RL) holds promise for the automated control of satellite orbits and the avoidance of spatial obstacles due to its ability to learn and adapt to dynamic environments through interaction, enabling autonomous, real-time decision-making for orbit adjustments. Satellite orbit control tasks manifest strong nonlinear characteristics, for which algorithms grounded on neural networks are more adept at handling. However, RL methods based on neural networks, which are highly nonconvex and nonsmooth functions, sometimes yield extreme action values that are not conducive to the orbit-constrained task at hand. In addition, in practical satellite task environments, limited observability is often faced. This limited observability not only increases the complexity of the algorithm but also affects the efficiency of identifying appropriate actions. To address this issue, we introduce a hierarchical parameterization mechanism to accelerate the search for appropriate actions. Hence, we introduce the integration of selfattention mechanisms with the Kalman filter into RL frameworks to accommodate scenarios with limited observability. Experimental results demonstrate a marked increase in the success rate of automated evasion maneuvers in environments with limited observability and orbital constraints, when facing pursuing satellites. Then, we performed a comparative analysis, applying Lambert guidance and RL-based guidance as control strategies for powerful confinement environment to validate our algorithm’s robustness.
UR - https://www.scopus.com/pages/publications/105048961887
U2 - 10.34133/space.0501
DO - 10.34133/space.0501
M3 - 文章
AN - SCOPUS:105048961887
SN - 2692-7659
VL - 6
JO - Space: Science and Technology (United States)
JF - Space: Science and Technology (United States)
M1 - 0501
ER -