Onboard Artificial Intelligence and Edge Computing for Spacecraft - a Review

Abstract

Spacecrafts gather more data than they can send back to Earth. When a satellite is in low earth orbit, it can be in contact with a ground station for just a few minutes during one orbit while constantly collecting data from its instruments, and a large portion of the data gathered is wasted due to cloud obstruction or lack of information. Edge computing in orbit shifts this paradigm by performing inference onboard, where only derived results will be sent back. This paper presents an overview of this topic starting from its first flight in 2020 until the current ongoing projects for platform-scale orbital compute infrastructure. Literature on this topic is classified into four categories: the hardware used and its radiation requirements, methods for compression of neural networks in order to fit in single-digit-watt power budget, fault-tolerance of systems under single event upsets, and distributed computation in space. The review ends with a discussion of the open issues that can be tackled by computer science practitioners. Special mention is made about the intersection of quantization and radiation-corruption. Fault-injection research has been performed on quantized neural networks on terrestrial edge devices, but no such study has yet been conducted for earth observation models based on orbital upset rates.