NAVI-Orbital Runs Gemma 3 Zero-Shot VLM Onboard LEO Satellite for Autonomous Scene Classification on April 16 2026

The system deployed a local vision-language model orchestrated by LangGraph agents on satellite-class edge hardware. It classified scenes, generated natural-language descriptions, and accepted plain-English re-tasking prompts instead of traditional command sequences. Ground tests, Flatsat validation, and live orbital captures confirmed end-to-end autonomy without downlink of raw pixels.

Data shows semantic compression inverted the conventional acquire-then-downlink pipeline. The 88.16 percent AID score held across previously unseen Earth imagery with hardware-accelerated GPU inference. This matches patterns in terrestrial edge VLM deployments but extends them to radiation-tolerant LEO compute where prior attempts relied on heavily quantized CNNs or required ground-in-the-loop confirmation.

Operationally, the demonstration removes downlink latency for time-critical observations such as disaster monitoring or vessel tracking. It shifts mission design from bandwidth-constrained raw data return toward selective transmission of text embeddings and dialogue outputs. Related work on orbital CNN accelerators and recent Gemma-3 efficiency benchmarks indicates the approach generalizes to additional 3U-6U platforms within existing power and thermal envelopes.

Next steps center on scaling agent coordination across constellations and validating radiation-induced error rates over longer durations.