Towards autonomous collaboration in heterogeneous satellite systems using machine learning.
Federated Satellite Systems (FSS), a type of Distributed Satellite Systems (DSS), enable heterogeneous satellites to collaborate and share unused resources—such as downlink opportunities, onboard storage, processing time, and even intermittent nodes for end-to-end communication. This cooperation can enhance the performance of existing missions and support the creation of virtual missions, allowing a better adaptation to the growing demands and requirements without the need to deploy additional constellations (similar to cloud computing).
The primary challenge of FSS lies in enabling efficient collaboration among these heterogeneous satellites, which are operated by different providers and experience intermittent communication windows. Traditional routing protocols struggle in this environment due to high convergence times, limited adaptability to dynamic topologies and traffic conditions, Earth dependencies, and stakeholders coordination efforts.
In contrast, Machine Learning (ML)-based routing protocols are gaining interest in this context as an alternative to support autonomous decisions. Although they may yield suboptimal solutions, they are inherently more scalable, adaptive, and cost-effective. Specifically, Reinforcement Learning (RL)-based routing protocols for LEO networks has shown to outperform traditional methods. This work proposes to extend RL-based routing strategies to the FSS paradigm, enabling more autonomous and cost-efficient decision-making across a dynamic network topology with intermittent inter-satellite links.
To contribute to the definition of a routing protocol to be applied in the context of FSS, characterized by a highly dynamic and heterogeneous environment, with resource-constrained nodes. To achieve this goal, the thesis is structured around the following objectives:
