Reinforcement Learning-Based Autonomous Control of a Spacecraft in the Solar Gravitational Lens’ Focus

A study addresses the challenge of autonomous spacecraft control near the Sun’s gravitational lens focal line via reinforcement learning techniques. The investigation examines the characteristic velocity requirements for targeting the focal line of a distant extended source, along with the precision of targeting and control function performance. Analysis is conducted under different observation scenarios: ideal spacecraft state, state with noise, and Einstein ring image. The research compares the effectiveness of different control approaches, specifically contrasting recurrent neural networks against fully connected layers that process stacked measurement inputs. Rigorous probability estimation for successful targeting is performed using Hoeffding’s inequality. Meta-reinforcement learning techniques applied to handle with maneuver execution errors.