Machine Learning for Spacecraft Dynamics and Control

The motion of a spacecraft in response to known disturbances can be predicted accurately using a well-defined dynamics model. Historically, most dynamics models have been well understood because spacecraft have been designed with simple geometries and limited flexibility. Even for more complex spacecraft, detailed analysis is often performed prior to launch to provide a suitable dynamics model for flight operations. Recently, however, interest in satellite servicing, in-space assembly, and debris removal has introduced new problems for controlling spacecraft in more uncertain scenarios. When a spacecraft becomes fixed to objects with unknown or uncertain properties, it becomes necessary to learn system dynamics through observation. For this project, the team is developing a new approach for learning spacecraft dynamics in a nonparametric way using gaussian process regression techniques. This approach allows previously unmodeled dynamics to be learned onboard a spacecraft with an uncertain or time-varying mass distribution, geometry, or control authority.

Image: Bryce Doerr / MIT