Bridging Reinforcement Learning and Online Learning for Spacecraft Attitude Control
Jacob G. Elkins, Rohan Sood, Clemens Rumpf
Abstract
Artificial intelligence is expected to revolutionize all areas of space operations in the coming years. The most advanced space systems will possess the ability to adapt and improve performance over time, or online learning. This work presents a novel framework that uses the highly researched artificial intelligence paradigm, reinforcement learning, to perform online learning. The spacecraft attitude control problem is used as a benchmark, with experimental results for using reinforcement learning to train neural spacecraft attitude controllers. Additionally, experimental results in a simulation environment are also shown to compare and contrast two state-of-the-art single-agent continuous control reinforcement learning algorithms to motivate their use in the online learning scenario.