Victor Vergara, AeroVironment; Francisco Viramontes, COSMIAC Research Center; Michael Felix, COSMIAC Research Center; Amanda Fetzer, AFRL; Windy Slater, AFRL
https://doi.org/10.64861/ABEM4763
Keywords: Edge AI, Space Debris, YOLO, Object Detection, Benchmark
Abstract:
Kessler Syndrome is a problem associated with space domain awareness (SDA) that will require the use of novel technologies to ameliorate the consequences of excessive space debris. The rise of artificial intelligence (AI) shows promise as a vital technology to identify debris, but on-orbit deployment presents a challenge for the space industry. One important problem is on-orbit data processing without the need for terrestrial communications or cloud computing services. Edge AI (EAI) computing is a recent innovation aiming to solve on-orbit computing problems where data collection and compute power are co-located. The main challenges of EAI consist of computing performance limitations imposed by the by hardware’s size, weight and power (SWaP). Some EAI frameworks implement AI models using Graphics Processing Unit (GPU) hardware with SWaP characteristics that allowed flight heritage. One example of this approach is the GPU-based NVIDIA Jetson Orin processor that has been integrated into several space missions. Pushing EAI computing towards lower SWaP characteristics gave rise to novel frameworks that deviate from the more typical Von Neumann Architecture. Analog computing, a concept that has existed since the beginning of electrical and computer engineering, can now provide EAI applications with high energy efficiency and massive parallelism. Neuromorphic, brain-inspired, technologies are also contributing to the race for lower SWaP. The Mythic Analog Matrix Processor (AMP) is an example of this class of devices that combines analog and neuromorphic concepts to accelerate AI computations at considerably lower power compared to digital counterparts.
Our motivation is to benchmark EAI hardware applied to a space debris detection task to compare their capabilities with respect to performance and energy consumption. The analog computing approach was tested using the Mythic AMP chip which promised large improvements in energy, cost, and performance efficiency. This device provides analog matrix processing by performing multiplication operations without the need for external memory scalability. Digital AI was implemented in an NVIDIA Jetson Orin processor, an example of a low-power, high-performance GPU integrated computer that was small enough for CubeSat missions. These systems can both run version five of the You-Only-Look-Once (YOLO) AI algorithm to perform object detection tasks with low power and high performance. While YOLO has been used in satellite imagery for intelligence, surveillance, and reconnaissance, this work explores EAI YOLO deployment for the debris detection SDA task. A simulated debris tracking scenario was constructed using the Computational Omni-Analysis Simulation Tool physics-based simulation software. A simulated optical sensor provided images of space debris in a low earth orbit that may be on a conjunction trajectory with the space mission system carrying the sensor. Building on previous work employing power-hungry GPU devices, we deployed AI inference into selected Edge AI devices to more closely represent the performance of on-orbit deployment.
Benchmarks were selected to be hardware agnostic given the notable differences between analog and digital processes. Performance metrics precision, recall, and mean average precision were chosen because they measure performance of any object detection task regardless of the specific AI algorithm. The other metric of interest is joules per inference measuring the amount of energy expended on each debris detection instance. Results show a better AI performance for digital hardware compared to analog, but at the cost of higher energy consumption when running the simulated SDA scenario. Our results also highlight the potential of in laboratorium benchmarking of Edge AI-SDA systems. Methods and procedures in this work focus on performing analysis and prediction, providing valuable insights for the SDA community.
We thank the AFRL Information and Spectrum Warfare Directorate for providing guidance support to make this possible. We also thank the AFRL Space Warfare Directorate’s SPACER laboratory for providing infrastructure to host the equipment necessary for data analysis. The views expressed are those of the author and do not necessarily reflect the official policy or position of the Department of the Air Force, the Department of Defense, or the U.S. government. Distribution Statement A: Approved for Public Release. Distribution is Unlimited. Public Affairs Release Approval # AFRL-2026-3094.
Date of Conference: September 15-18, 2026
Track: SDA Systems & Instrumentation