Multi-Agent Reinforcement Learning Empower Space Unmanned Systems: Methods, Challenges and Opportunities

With the advancement of space technology towards intelligence and clusterisation, unmanned space systems demonstrate immense potential in strategic areas such as deep space exploration and Earth observation. However, traditional centralized control paradigms face significant challenges in adressing highly dynamic environments, distributed tasks, and strict resource constraints. Leveraging its distributed decision-making architecture and co-evolutionary mechanisms, multi-agent reinforcement learning (MARL) offers a breakthrough solution for building autonomous and resilient intelligent space systems. This paper systematically explores MARL’s technological empowerment pathways, methodologies, engineering challenges, and opportunities in unmanned space systems. It analyzes the technical bottlenecks in core scenarios (e.g., collaborative communication for satellite clusters, multi-spacecraft control). Moreover, It reveals the application mechanisms of MARL in critical domains, including dynamic spectrum allocation, on-board edge computing, and robust collaborative control. Finally, the paper proposes an integrated collaborative intelligence architecture that incorporates space-dynamics constraints with innovative MARL algorithms. This framework aims to drive the evolution of space systems toward a new paradigm of autonomous decision-making, resilient anti-jamming capabilities, and efficient collaboration. This research seeks to provide theoretical support and a technological roadmap for the next-generation space-based intelligent networks.