Distributed inferencing and federated learning, for distributed-edge-AI miniaturised satellite constellations
Abstract
Although orbital edge computing is in its infancy, research in the literature shows promising results in bringing artificial intelligence to the space edge for optimizing missions, providing actionable insights in near-real time, addressing the limitations of classical communication pipelines, and ensuring data privacy. While a large collection of satellites employing pre-trained ML models would represent a multiplication of capability, this would be no more than the sum of the parts. To improve the training for such a collection, the data from each satellite would need to be downlinked and aggregated on the ground, and scalability would be constrained. A far more powerful approach is for the collection of satellites itself to be an intelligent complex system, where each satellite is capable of on-orbit training that can be shared and combined across the constellation to build a better model than each individual component. Research into edge-AI constellation approaches such as distributed inferencing and federated learning, and the scheduling challenges that arise with such approaches, is required. Performing the research and development required to address this challenge and deliver robust, scalable technologies for intelligent constellations cannot be achieved without testing and validating the approaches and their scalability on constellations. For this reason, a networked space-edge-AI test bed is required that incorporates actual edge-AI hardware and potentially on-orbit satellites, yet can be scaled with the inclusion of large numbers of virtual satellites. In a collaboration between UNSW Canberra Space and AFRL RI, the proposed research seeks to achieve: the development of a lab-based, modular and scalable, virtual and hardware-in-the-loop (HIL) test bed for accelerating AI-enabled-miniaturised-satellite-constellation R and D; and development and demonstration on the test bed, of methods for distributed inferencing and federated learning on AI-enabled-miniaturised-satellite-constellations.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jan 21, 2022
- Source ID
- FA23862114047XX0
Entities
People
- Russell Boyce
Organizations
- Air Force Office of Scientific Research
- United States Air Force
- University of New South Wales