Aerospace Control and Guidance Systems Committee

Announcements

You must first log in to access prior meeting presentations, register for a meeting, or nominate some for the Ward Award.

If you do not have a login account, or cannot remember the email address associated with your account, please click on the Application Form link below.

Login

Site Search

Search our site:

Upcoming Events

Photos

Meeting Highlights New!

Subcommittee S

Prior Meetings

Abstracts may be viewed by anyone. Presentations are only available to active members who have logged in.

Meeting 133
(coming soon)

Meeting 132
(coming soon)

Home

Ward Memorial Award

Planning Advisory Board

Downloads

Constitution and By-Laws

About

History

Contact Us

← Return to agenda

Meeting	ACGS Committee Meeting 126 - Virtual - March 2021
Agenda Location	7 SUBCOMMITTEE C – AVIONICS AND SYSTEM INTEGRATION 7.3 Future of Aerospace Autonomy: How to Integrate Machine Learning with Guidance, Navigation, and Control
Title	Future of Aerospace Autonomy: How to Integrate Machine Learning with Guidance, Navigation, and Control
Presenter	Soon-Jo Chung
Affiliation	Caltech/JPL
Available Downloads*	presentation
	Downloads are available to members who are logged in and either Active* or attended this meeting.
Abstract	One common theme of our research projects at Caltech’s Center for Autonomous Systems and Technologies (CAST) and my research group http://aerospacerobotics.caltech.edu/ is to systematically leverage AI and Machine Learning (ML) towards achieving safe and stable autonomy of robotic and aerospace systems, such spacecraft swarms and drones. Stability and safety are often research problems of control theory and robotics, while conventional black-box AI approaches lack the much-needed robustness, scalability, and interpretability, which are indispensable to designing control and autonomy engines for safe-critical aerospace robotic systems. I will show how to apply spectral normalization (SN) to constrain the Lipschitz constant of Deep Neural Networks (DNN). Leveraging this Lipschitz property, we design a nonlinear tracking controller using the learned model and prove system stability with disturbance rejection with last-layer adaption for real-time learning. Recent results on neural-network-based contraction metrics (NCMs) and its stochastic extensions (NSCM) for safe motion planning will be presented. A Neural Contraction Metric (NCM) is a neural network model of an optimal contraction metric constructed to minimize an upper bound of the steady-state Euclidean distance between perturbed and unperturbed system trajectories. The NCM framework is shown to be applicable to robust and adaptive control and estimation of nonlinear systems with bounded disturbances, stochastic perturbation, and parametric uncertainty. I will also briefly introduce the method of ML-based Global-to-Local Autonomy Synthesis (GLAS) for multi-vehicle planning that avoids local minima while maintaining high scalability and computational efficiency.