Deep Reinforcement Learning for On-Orbit Servicing Scheduling for Low and High Thrust Propulsion
Year: 2026
Type: dissertation
Abstract: The increasing density and economic value of satellite infrastructures make on-orbit servicing (OOS) an important capability for future space operations. This thesis presents the Integrated On-Orbit Services Scheduling Optimization (IOOSSO) framework, a deep reinforcement learning approach for dynamically scheduling OOS missions in multi-orbit environments. The problem is formulated as an event-driven semi-Markov decision process with stochastic service needs, orbital refuelling depots, and fuel, time, and orbital-transfer constraints. High-thrust (HT) and low-thrust (LT) architectures are modeled within a continuous orbital representation, with feasible rendezvous transfers constructed on the fly. A Maskable Proximal Policy Optimization agent learns servicing policies that maximise an integrated-system net present value (NPV) by balancing preserved fleet revenue against infrastructure, propellant, and asset-loss costs. Medium Earth Orbit case studies show that servicing profitability is strongly horizon- and demand-dependent. Neither architecture recovers its infrastructure cost over one year, whereas both break even within the first year of a five-year campaign and create approximately \$500M of value relative to a replace-on-failure baseline. LT is favored at low demand because of its propellant efficiency, but HT overtakes it at about 34 satellites as transfer speed becomes decisive. Additional depots create value only when constellation geometry makes inter-orbit manoeuvres expensive, and optimized placement can nearly substitute for depot count. Finally, reinforcement learning provides decisive gains for LT and under constrained refuelling logistics, while a greedy heuristic remains competitive for HT when transfers are short and depot access is benign.
Author: claret, nathan
Institution: Concordia University
Language: English
Subfield: Aerospace Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7221486360