NAVI-Orbital Demonstrates First Onboard Zero-Shot Vision-Language Model on Low Earth Orbit Satellite
Researchers have demonstrated NAVI-Orbital, the first onboard software system to execute a vision-language model directly on a Low Earth Orbit spacecraft. Operating a local Gemma 3 model orchestrated by LangGraph, the system performs real-time semantic compression and supports natural-language command execution.
The rapid growth of Earth observation data has created a severe bottleneck due to limited downlink bandwidth and the manual overhead of ground-based processing. To address this, NAVI-Orbital shifts the processing paradigm by executing a vision-language model, specifically Gemma 3, directly on satellite-class edge computers. This architecture enables real-time scene classification, natural-language description generation, and interactive operator dialogue onboard the spacecraft. Orchestrated by a LangGraph-based state machine, the system coordinates dedicated agent workflows for object detection and dialogue. Operators can re-task the satellite using plain-English prompts instead of traditional command sequences. This edge capability allows the system to compress raw image data semantically, prioritizing critical information for downlink and drastically reducing the required transmission bandwidth. Validation of the system included ground benchmarking on the AID dataset, achieving an 88.16% accuracy rate, followed by successful live in-orbit execution using uncorrected imagery. The demonstration, conducted on April 16, 2026, proves the feasibility of running modern foundation models on edge-compute hardware in space without requiring instrument-specific fine-tuning.
Primary source
Open arXiv ↗This article summarizes the source. Check the original for full details and conditions.
Comparison
| Aspect | Before / Alternative | After / This |
|---|---|---|
| Data Processing Location | Ground stations after transferring raw, high-bandwidth imagery downlinks | Onboard edge computing to perform semantic compression prior to downlink |
| Satellite Re-tasking | Rigid, conventional hardware command sequences | Plain-English natural-language prompts via agent workflows |
| System Orchestration | Procedural embedded flight software routines | Graph-based state machine using LangGraph to coordinate agents |
| Model Adaptability | Highly specialized models fine-tuned for specific flight instruments | Zero-shot foundation model (Gemma 3) requiring no instrument-specific fine-tuning |