Session
Weekend Session V: Automation - Research and Academia
Location
Utah State University, Logan, UT
Abstract
Despite their potential for planetary missions, CubeSats have significant constraints due to their limited size, weight, and power. For lunar surface exploration, for example, collecting high resolution data requires flying as low as possible to get adequately close to the target. We propose a machine learning approach to orbital maintenance manoeuvres in very low lunar orbits using reinforcement learning (RL) to train an onboard neural network (ONN). This approach determines orbital manoeuvres based on the current state of the spacecraft and locally determinable knowledge of the lunar environment. The ONN seeks to maximise the mission duration while maintaining a target altitude. The algorithm was developed using popular open-source deep learning libraries and was trained and evaluated using GMAT (General Mission Analysis Tool) as the simulation environment.
Onboard sensors play a crucial role in this approach, as they allow the spacecraft to make real-time decisions based on current knowledge of the lunar environment and respond to unexpected changes. This is particularly important at extremely low altitudes, where limited access to ground-based control can threaten the feasibility of a mission. With the growing potential for onboard neural networks in space exploration, this approach seeks to contribute to the current leap forward in autonomous spacecraft control.
Our algorithm is designed to navigate the spacecraft safely in challenging conditions, such as those presented by the lunar gravity field which destabilizes orbits and creates the risk of unscheduled high velocity landings. The focus is on controlling altitude to avoid collisions with topographical features, not on targeting specific ground features, and specifically in situations where orbit determination is unavailable. By taking these unique challenges into consideration, we aim to improve the viability of lunar terrain flying.
How Low Can You Go? An AI-Powered Solution for Autonomous Spacecraft Control in Very Low Lunar Orbits
Utah State University, Logan, UT
Despite their potential for planetary missions, CubeSats have significant constraints due to their limited size, weight, and power. For lunar surface exploration, for example, collecting high resolution data requires flying as low as possible to get adequately close to the target. We propose a machine learning approach to orbital maintenance manoeuvres in very low lunar orbits using reinforcement learning (RL) to train an onboard neural network (ONN). This approach determines orbital manoeuvres based on the current state of the spacecraft and locally determinable knowledge of the lunar environment. The ONN seeks to maximise the mission duration while maintaining a target altitude. The algorithm was developed using popular open-source deep learning libraries and was trained and evaluated using GMAT (General Mission Analysis Tool) as the simulation environment.
Onboard sensors play a crucial role in this approach, as they allow the spacecraft to make real-time decisions based on current knowledge of the lunar environment and respond to unexpected changes. This is particularly important at extremely low altitudes, where limited access to ground-based control can threaten the feasibility of a mission. With the growing potential for onboard neural networks in space exploration, this approach seeks to contribute to the current leap forward in autonomous spacecraft control.
Our algorithm is designed to navigate the spacecraft safely in challenging conditions, such as those presented by the lunar gravity field which destabilizes orbits and creates the risk of unscheduled high velocity landings. The focus is on controlling altitude to avoid collisions with topographical features, not on targeting specific ground features, and specifically in situations where orbit determination is unavailable. By taking these unique challenges into consideration, we aim to improve the viability of lunar terrain flying.
