A reinforcement learning training and testing framework for aerospace missions

With the increasing complexity of aerospace missions, the limitations of traditional task planning methods in dynamic and uncertain environments have become increasingly apparent. Reinforcement learning, as an autonomous decision-making optimization technique, offers a new approach to address this challenge. However, its application in the aerospace domain still faces difficulties in environment modeling, testing, and verification. This paper proposes a reinforcement learning training and testing framework specifically designed for aerospace missions. The framework constructs a multi-task meta-model integrated environment based on the SpaceSim simulation platform, designs a unified interface specification compliant with the Markov decision process, and establishes a modular and visual training and testing platform. The effectiveness of the framework is validated through four typical mission scenarios: orbital interception, maneuver avoidance, remote sensing-maneuver collaboration, and constellation mission planning. Experimental results demonstrate that the framework not only supports the training and testing of different reinforcement learning algorithms but also outperforms traditional optimization methods, providing a customizable, extensible, and practical platform for addressing intelligent decision-making problems in aerospace missions.