Session
Frank J. Redd Student Competition
Location
Salt Palace Convention Center, Salt Lake City, UT
Abstract
As the number of Space Vehicles (SVs) orbiting Earth grows, downlink bitrate bottlenecks necessitate a shift to Orbital Edge Computing (OEC). Size, Weight, and Power, and Cost (SWaP-C) limit onboard Machine Learning (ML) autonomy, and State-of-the-Art (SotA) systems deploy static ML models that do not adapt to dynamic orbit conditions. This work proposes and evaluates RADMA, a resource-aware framework that selects Neural Network (NN) models in response to changing latency and energy constraints. We quantify the performance of RADMA on an open-hardware picosatellite payload that hosts a Google Coral Edge Tensor Processing Unit (TPU). Orbit simulation and modeling predict processing deadlines and satellite energy, allowing an onboard scheduler to select from a set of pre-characterized Convolutional Neural Networks (CNNs) spanning different widths and depths. We evaluate RADMA with simulated case studies spanning elliptical Low Earth Orbit (eLEO) and polar orbit. Compared to static deployment baselines, RADMA increases the number of correct inferences by up to 51% over high-speed models and outperforms maximum-accuracy deployments by over 900%. By predicting energy consumption, RADMA virtually eliminates unplanned downtime, demonstrating the value of dynamic ML model deployment on resource-constrained edge Domain-Specific Accelerators (DSAs) in space.
Document Type
Event
RADMA: Resource-Aware, Dynamic Deployment of ML Architectures on an Edge TPU in Space
Salt Palace Convention Center, Salt Lake City, UT
As the number of Space Vehicles (SVs) orbiting Earth grows, downlink bitrate bottlenecks necessitate a shift to Orbital Edge Computing (OEC). Size, Weight, and Power, and Cost (SWaP-C) limit onboard Machine Learning (ML) autonomy, and State-of-the-Art (SotA) systems deploy static ML models that do not adapt to dynamic orbit conditions. This work proposes and evaluates RADMA, a resource-aware framework that selects Neural Network (NN) models in response to changing latency and energy constraints. We quantify the performance of RADMA on an open-hardware picosatellite payload that hosts a Google Coral Edge Tensor Processing Unit (TPU). Orbit simulation and modeling predict processing deadlines and satellite energy, allowing an onboard scheduler to select from a set of pre-characterized Convolutional Neural Networks (CNNs) spanning different widths and depths. We evaluate RADMA with simulated case studies spanning elliptical Low Earth Orbit (eLEO) and polar orbit. Compared to static deployment baselines, RADMA increases the number of correct inferences by up to 51% over high-speed models and outperforms maximum-accuracy deployments by over 900%. By predicting energy consumption, RADMA virtually eliminates unplanned downtime, demonstrating the value of dynamic ML model deployment on resource-constrained edge Domain-Specific Accelerators (DSAs) in space.
