Spacecraft guidance, navigation, and control systems need to operate under substantial uncertainties. This presents challenges when designing these control systems using conventional methods that require a certain level of knowledge of the system being
controlled. Intelligent control systems have emerged as a means of addressing these challenges. These systems combine theories from automatic control, operations research, and artificial intelligence to derive controllers that can deal with different types
of uncertainty. Various methods from the field of artificial intelligence can be used to develop intelligent control systems, however these are often computationally expensive which limits their applicability in spacecraft control problems. A key feature of intelligent control is the ability to adapt the control system online, which presents further
difficulties when this must be done onboard a spacecraft. This thesis explores the use of reinforcement learning techniques for intelligent control applied to spacecraft powered descent. The proposed approach combines a reinforcement learning agent for handling uncertainties with conventional optimisation methods to improve the agent’s performance. In addition, the agent updates its control policy online using a novel update mechanism called Extreme Q-Learning Machine, which allows the control system to
operate in a changing environment. To demonstrate the potential for this method to be implemented onboard spacecraft, results are shown from running online updates on flight suitable hardware. This work provides one possible avenue for increasing the level of intelligence of spacecraft control systems.
| Date of Award | 24 Apr 2026 |
|---|
| Original language | English |
|---|
| Awarding Institution | - University Of Strathclyde
|
|---|
| Sponsors | University of Strathclyde |
|---|
| Supervisor | Annalisa Riccardi (Supervisor) & Edmondo Minisci (Supervisor) |
|---|