Abstract
Spacecraft trajectory design is critical to successful space missions, particularly with the increasing number of spacecraft and mission complexity. Satellite constellation deployment, reconfiguration, orbit raising and deep-space missions all depend on effective trajectory design. In addition, spacecraft autonomy is a major barrier to increasing the scope, ambition and affordability of both Earth-based and deep-space missions. The current state-of-the-art in spacecraft operations is still to guide space missions from the ground with extensive human intervention. Whilst automated guidance, planning and trajectory design tools do exist, they often lack the vital skill of human operators, who can act under environmental and mission uncertainty.
The goal of this thesis was to develop and investigate a lightweight and closed-loop control law that can be used for both initial trajectory design and subsequent on-board guidance. The motivation behind this research is to combine the stable yet sub-optimal nature of Lyapunov control laws with the exploration and state-dependence offered by reinforcement learning techniques. This has resulted in the development of a novel Reinforced Lyapunov Controller. The Lyapunov stability implications are examined and an analytical expression for the state-weight Jacobian is presented. Performance in Keplerian dynamics is investigated to assess the optimality and stability of the approach. New training procedures in the presence of unmodelled dynamics including perturbations, eclipse events and stochastic errors are presented. A cone-clock angle approach is devised to explore the additional degrees of freedom and further ensure Lyapunov stability. Results show the Reinforced Lyapunov Controller is optimal and stable in modelled dynamics, and robust to uncertain and stochastic environments. Including such uncertainty in the training procedure can further improve the performance. The versatility is considered by approximating finite-burn trajectories and incorporating operational constraints. Finally, the potential of the Reinforced Lyapunov Controller for Earth-Moon spiral transfers is investigated, exploring the use of a two-body control law in a three-body environment. In both cases, the Reinforced Lyapunov Controller is able to compete with conventional evolutionary algorithm methods.