Session

Science/Mission Payloads

Location

Salt Palace Convention Center, Salt Lake City, UT

Abstract

The rapid maturation of multimodal vision-language models (VLMs) has significantly expanded artificial intelligence applications at the edge, including spaceborne systems. In this work, we present one of the first demonstrations of deploying a compact vision-language model directly on a very-high-resolution Earth-observation satellite payload processor, enabling autonomous on-orbit scene understanding and near real-time decision-making. This represents a paradigm shift in satellite operations, moving from ground-centric processing toward intelligent, self-directed spacecraft. 

Conventional Earth-observation missions rely on ground-based image processing and interpretation, introducing latency, downlink bandwidth constraints, and limited operational responsiveness. By embedding a VLM on board the satellite, visual reasoning can be performed in situ, allowing the spacecraft to autonomously interpret imagery, generate semantic descriptions, and prioritize data for downlink based on mission-relevant features without continuous ground intervention. 

To validate this concept, we integrated Google's Gemma-3n, a lightweight multimodal vision-language model, onto an NVIDIA Jetson Orin-based payload processing system within a proprietary, ruggedized, very-high-resolution satellite computer architecture. We demonstrate that 200 TOPS AI performance can be achieved under space-relevant power and thermal constraints. Sufficient computational margin was verified to support inference on the Gemma-3n model (~3.8B parameters) while maintaining total power consumption compatible with constrained on-board resources (~15W). Through custom quantization and model-level optimizations, end-to-end inference latencies below two seconds were achieved on satellite imagery acquired in orbit. 

This work substantiates the feasibility of spaceborne edge multimodal AI beyond single-task CNN pipelines, and provides a practical path toward autonomous constellation operations such as rapid response to emergent events, cooperative inter-satellite tasking, and resilient on-orbit intelligence when ground connectivity is constrained.

Document Type

Event

Available for download on Saturday, August 22, 2026

Share

COinS
 
Aug 25th, 5:00 PM

Payload Operation Using On-Board Vision Language Model

Salt Palace Convention Center, Salt Lake City, UT

The rapid maturation of multimodal vision-language models (VLMs) has significantly expanded artificial intelligence applications at the edge, including spaceborne systems. In this work, we present one of the first demonstrations of deploying a compact vision-language model directly on a very-high-resolution Earth-observation satellite payload processor, enabling autonomous on-orbit scene understanding and near real-time decision-making. This represents a paradigm shift in satellite operations, moving from ground-centric processing toward intelligent, self-directed spacecraft. 

Conventional Earth-observation missions rely on ground-based image processing and interpretation, introducing latency, downlink bandwidth constraints, and limited operational responsiveness. By embedding a VLM on board the satellite, visual reasoning can be performed in situ, allowing the spacecraft to autonomously interpret imagery, generate semantic descriptions, and prioritize data for downlink based on mission-relevant features without continuous ground intervention. 

To validate this concept, we integrated Google's Gemma-3n, a lightweight multimodal vision-language model, onto an NVIDIA Jetson Orin-based payload processing system within a proprietary, ruggedized, very-high-resolution satellite computer architecture. We demonstrate that 200 TOPS AI performance can be achieved under space-relevant power and thermal constraints. Sufficient computational margin was verified to support inference on the Gemma-3n model (~3.8B parameters) while maintaining total power consumption compatible with constrained on-board resources (~15W). Through custom quantization and model-level optimizations, end-to-end inference latencies below two seconds were achieved on satellite imagery acquired in orbit. 

This work substantiates the feasibility of spaceborne edge multimodal AI beyond single-task CNN pipelines, and provides a practical path toward autonomous constellation operations such as rapid response to emergent events, cooperative inter-satellite tasking, and resilient on-orbit intelligence when ground connectivity is constrained.