NASA Goddard Space Flight Center (GSFC) has developed a CubeSat-sized Tensor Processing Unit (TPU) co-processor card for accelerating the inference of Artificial Intelligence (AI) models known as SpaceCube Low-power Edge AI Resilient Node (SC-LEARN), which has recently launched and is now successfully integrated on the International Space Station as part of the Space Test Program - Houston 9 - SpaceCube Edge-Node Intelligent Collaboration (SCENIC). Although SC-LEARN enables the practical in-flight application of AI for the first time for GSFC, the use of deep learning in space has remained severely limited due to various factors such as the lack of compatible TPU models, flight software and software interfaces, formal testing and validation procedures, and tools for model explainability and decision verification most of all. This work seeks to address these challenges to bridge the gap between academic state-of-the-art AI models and their application in spaceflight, with an initial focus on developing visual perception models that can directly enhance scene understanding and situational awareness within autonomous systems.
More »The maturity of space-grade deep learning has yet to reach a level acceptable for mission-critical operation. This can be attributed to the lack of standardized verification tools at NASA, where a cohesive methodology for developing and testing AI systems is required to produce more reliable and trustworthy models at a faster pace. To that end, this work seeks to lay the foundation for mission-agnostic AI compliance through the Goddard AI Verification and INtegration (GAVIN) tool suite. GAVIN represents a unification of common AI development practices under one software repository, with an emphasis on exhaustive verification and explainability procedures. This proposal explores the use of such procedures to develop space-grade visual perception models that directly enhance situational awareness capabilities. The key contributions of this proposal include:
1. Creation of the GAVIN tool suite, consisting of four software modules and a NOS3 plugin that enables photo-realistic simulation for model training and evaluation.
2. Production of four vision-based deep learning models for planetary and small body terrain detection, tracking, and identification using GAVIN.
3. Development of three culminating demonstrations that introduce advanced visual perception capabilities for low-illumination lunar Terrain Relative Navigation (TRN), hazard detection and tracking, and hyperspectral image reasoning.
Organizations Performing Work | Role | Type | Location |
---|---|---|---|
Goddard Space Flight Center (GSFC) | Lead Organization | NASA Center | Greenbelt, Maryland |