Spacecraft Attitude Control Under Multiple Constraints: A Progressive Deep Reinforcement Learning Approach

Abstract

An end-to-end large-angle attitude tracking control framework for spacecraft under multiple physical constraints based on deep reinforcement learning (DRL) is proposed, including attitude pointing constraints, moment amplitude constraints and internal disturbances. Unlike conventional Euler angle and quaternion representations, this methodology uses the rotation matrix on the special orthogonal group SO(3) to establish a global, singularity-free posture representation ideal for large-angle maneuvering. To satisfy high-precision tracking requirements, a phased training strategy combined with progressive neural network (PNN) has been introduced to achieve effective cross-task knowledge transfer, improve training efficiency, and prevent catastrophic forgetting. The training process is guided by a dynamic weighted composite reward function, which automatically balances long-term mission objectives and immediate performance feedback. This design ensures strict compliance with pointing constraints, while compensating for the adverse effects of moment constraints and internal interference on control performance. Four comparative experiments within an SO(3) Gym environment demonstrate the overall effectiveness of the proposed algorithm and its superior performance over traditional analytical methods and baseline reinforcement learning (RL) approaches.