Reinforcement learning-based attitude control for a barbell electric sail.

Area of Science:

• Aerospace Engineering
• Control Systems
• Artificial Intelligence

Background:

• Electric solar wind sails (E-sails) offer propellant-free propulsion but present complex control challenges due to their nonlinear and under-actuated dynamics.
• Traditional control methods struggle with the intricate attitude control requirements of E-sail systems.

Purpose of the Study:

• To develop and evaluate a reinforcement learning (RL)-based control scheme for the attitude control of a barbell E-sail system.
• To address the limitations of conventional control schemes in managing E-sail systems.

Main Methods:

• A two-stage RL control strategy was designed, utilizing the Proximal Policy Optimization (PPO) algorithm and neural networks for policy emulation and updates.
• The system's attitude dynamics were modeled using a nonsingular formulation.
• Real-time attitude feedback control was achieved through a state-to-control output mapping derived from the learned RL strategy.

Main Results:

• The proposed RL-based control scheme successfully regulated tether voltage differences to achieve desired E-sail attitudes.
• Simulations demonstrated effective attitude adjustment capabilities of the RL controller.
• Comparisons with Nonlinear Model Predictive Control (NMPC) showed significant reductions in computation time without compromising control accuracy.

Conclusions:

• The developed RL-based control scheme provides an efficient and accurate solution for E-sail attitude control.
• This approach offers a promising alternative to conventional methods, particularly for complex, nonlinear systems.
• The study highlights the potential of RL in advancing propellant-free space propulsion technologies.