Neural Network Optimal Control in Astrodynamics: Application to the Missed Thrust Problem – Acta Astronautica

Note: This post is adapted from my journal paper, which was accepted to Acta Astronautica in Spring 2020. You can read the full paper here

Abstract – Neural Network Optimal Control in Astrodynamics

While high-efficiency propulsion techniques are enabling new mission concepts in deep space exploration, their limited thrust capabilities necessitate long thrusting arcs and make spacecraft more susceptible to missed thrust events  (MTE). To correct for such mishaps, most spacecraft require updated trajectories that are relayed from Earth. While this solution is viable for spacecraft near Earth, in deep space, where one-way communication time is measured in hours, a delay in transmission may prolong the time of flight or result in a complete loss of mission. Such problems can be alleviated by increasing the spacecraft’s onboard autonomy in guidance. This paper demonstrates how a computationally lightweight neural network can map the spacecraft’s state to a near-optimal control action, autonomously guiding a spacecraft within different astrodynamic regimes and optimality criteria. The neural network is trained using supervised learning and datasets comprised of optimal state-action pairs, as determined through traditional direct and indirect methods. Additionally, the neural network-designed solutions retain optimality and time of flight corresponding to traditional trajectories. Finally, the same neural networks can autonomously correct for most missed thrust events encountered on long-duration low-thrust trajectories. The presented results provide a path for mitigating risks associated with the use of high-efficiency low-thrust propulsion techniques.

Key Takeaways

The full paper is online here, but here are, in my opinion, the most interesting aspects

• A neural network can guide a spacecraft to its target >97% of the time
• The same network can autonomously repair MTEs in an interplanetary trajectory >77% of the time. (There were no MTEs in the training dataset)
• The neural network is smaller in size (about 1000) than neural networks already deployed in space missions

Want More Gereshes?

If you want to receive new Gereshes blog post directly to your email when they come out, you can sign up for that here!

Don’t want another email? That’s ok, Gereshes also has a twitter account and subreddit!