Using Reinforcement Learning to Design a Low-Thrust Approach into a Periodic Orbit in a Multi-Body System
Authors:
- 5 January 2020
- conference paper
- Published by American Institute of Aeronautics and Astronautics (AIAA)
Abstract
Exploring multi-body systems with low-thrust-enabled small satellites opens up a new frontier of scientific opportunities. However, the limited control authority of these spacecraft introduces significant challenges during trajectory and maneuver design. In this paper, a deep reinforcement learning method is used to design a control profile for a low-thrust-enabled small satellite to approach a periodic orbit from a nearby initial condition and over a short time horizon. Multiple neural networks are trained to recover controlled paths that meet a variety of objectives while tracking to a reference trajectory on a stable manifold. Monte Carlo simulations for three distinct formulations of a reward function are performed using a control scheme derived from trained neural networks.Keywords
This publication has 10 references indexed in Scilit:
- A low-thrust-enabled SmallSat heliophysics mission to Sun-Earth L52018
- Trajectory design for a cislunar CubeSat leveraging dynamical systems techniques: The Lunar IceCube missionActa Astronautica, 2018
- Jumping mechanisms of Trojan asteroids in the planar restricted three- and four-body problemsCelestial Mechanics and Dynamical Astronomy, 2015
- Role of Invariant Manifolds in Low-Thrust Trajectory DesignJournal of Guidance, Control, and Dynamics, 2009
- Fundamentals of Electric Propulsion2008
- REINFORCEMENT LEARNING AND DYNAMIC PROGRAMMING1995
- Learning Without State-Estimation in Partially Observable Markovian Decision Processes1994
- Function Optimization using Connectionist Reinforcement Learning AlgorithmsConnection Science, 1991
- Modifications of the Restricted Problem1967
- Steps toward Artificial IntelligenceProceedings of the IRE, 1961