A Multi-Agent Deep Reinforcement Learning Framework for the Stable Landing of a Flexibly Connected Three-Node Space Probe

Exploring asteroids in the solar system is crucial for human survival and development in the future, since the asteroids may contain important clues about the origin of life which can bring important inspiration to humanity. However, due to the weak gravitational force of asteroids, as well as the limited knowledge about them, it is challenging for a space probe to land on these asteroids' surface steadily. To this end, we propose a flexible connected three-node probe to significantly boost the landing stability, in which each node is relatively independent in behaviors but constrained with each other. As a multi-agent system with constraints, a cooperative behavior planning and decision-making method is very necessary. In this paper, we propose a multi-agent deep reinforcement learning (MADRL) framework for addressing the stable landing of a flexibly connected three-node space probe. In order to overcome the uncertainty of the environment, a training paradigm of centralized training decentralized execution (CTDE) is adopted, where the information exchange between agents is taken into account. Moreover, we integrate gated recurrent unit (GRU) modules into actor and critic networks to preserve the historical information, so that the learned strategy can apply to uncertain asteroid environments and implement stable probe landing. Experimental results demonstrate that the proposed method outperforms other reinforcement learning-based methods in terms of convergence and stability.