Sage Schaffer

Applying Reinforcement Learning to Optimize Lower-Earth Orbital Transfers

Started Sept. 11, 2023

Abstract or project description

DESCRIPTION: This project with be focused on developing a program that may be run on a satellite in order to determine the optimal design of a spacecraft servicing network. Building on two-body principals, first a Monte Carlo simulation will be used to select good initial states for the spacecraft. Then the same data will be used to train a neural net that can be deployed to the satellite in order to determine transfers in real time.

ABSTRACT:

This research develops a reinforcement learning (RL)-based neural network (NN) to optimize orbital transfers in low Earth orbit. Traditional methods, such as the Monte Carlo (MC) simulation, require numerous computational resources and iterations; the proposed RL is benchmarked by the MC and produces more accurate results across all metrics, past a threshold of training. With further development reducing necessary on-board computations, the proposed RL has potential viability to replace MC for satellite servicing missions. The RL was trained on four Hohmann and non-coplanar transfer scenarios, and the results demonstrate progression in the NN’s accuracy given more timesteps. The results indicate RL’s capability to predict optimal trajectories and adapt to varying scenarios, offering potential reductions in cost and computation for future satellite servicers. This proof of concept establishes the foundation for RL-based NN applications in more complex orbital mechanics problems, specifically real-time scenarios with live trajectory updates.