WASHINGTON — In a a push to bring cutting edge computing to the space sector, satellite infrastructure provider Loft Orbital has announced a landmark agreement with NASA’s Jet Propulsion Laboratory (JPL) to test advanced artificial intelligence directly on operational spacecraft.
The initiative is part of a NASA-funded project called Federated Autonomous Measurement (FAME).
Initial testing began earlier this month using a Loft Orbital satellite already in orbit, with a second phase of more advanced tests slated for 2027 and 2028 utilizing the company’s next-generation spacecraft.
Overcoming the Ground Bottleneck
The primary objective of the FAME project is to fully automate “tip-and-cue” operations; a method where a wide-field sensor detects an event and signals a high-resolution sensor to zoom in for closer inspection.
Currently, this process is hobbled by communication delays. Satellites must downlink massive, raw imagery files to ground stations, where terrestrial computers analyze the data before operators beam new targeting commands back up to a second satellite. By shifting the AI processing directly onto the spacecraft, the entire analysis happens in orbit.
Traditional vs. Autonomous Edge AI Processing
- Traditional Method: [Satellite Captures Image] ──> [Downlinks Massive Raw Data] ──> [Ground Team Analyzes Data] ──> [Uplinks New Command To Second Satellite]. This method takes hours/days
- Loft Orbital Edge AI: [Satellite Captures Image] ──> [Onboard AI Detects Event] ──> [Intersatellite Link Commands Nearby Satellite]. This method takes minutes
“You can use different assets with processing at the edge to capture, sense, understand and send insights about what’s going on on Earth without having to downlink big amounts of data,”
– Paul Lasserre, General Manager for AI at Loft Orbital
The Convergence of Hardware and Small AI Models
Executing high-performance machine learning in space presents fierce engineering hurdles, primarily due to the severe power, thermal, and weight constraints of satellite hardware. This mission was impossible until very recently. The breakthrough relies on a critical intersection of two technologies maturing simultaneously:
- Space-Qualified Accelerators: Onboard satellite processors capable of executing dense computational workloads in real time.
- Small Multimodal Models (SMMs): The emergence of compact, open-source AI models featuring multimodal reasoning. These models are highly optimized to run on low-power hardware while remaining trained on a vast corpus of data, allowing them to autonomously identify complex environmental features without explicit pre-programming.
“The constraint exists, but it’s not really limiting…We can run state-of-the-art models if we have prepared them well.”
– Paul Lasserre, General Manager for AI at Loft Orbital
Activating “Patrol Mode” for Global Response
By enabling real-time autonomy, Loft Orbital aims to unlock a continuous “patrol mode” for constellations.
A single, wide-angle sensor can remain perpetually active, using its local AI model to scan the Earth for anomalies. If the model flags a critical event, such as a flash wildfire ignition, a marine oil spill, or sudden military asset movements, it can instantly send targeted coordinates to neighboring satellites via intersatellite laser links for rapid, high-resolution follow-up.
Loft Orbital’s Future Altair Infrastructure
To fully scale this architecture, Loft Orbital is actively developing Altair, a dedicated constellation consisting of 10 satellites. The constellation will feature a robust, interconnected value chain
[Multi-Sensor Payloads] ──> [Onboard Edge AI Compute] ──> [Intersatellite Laser Links]
“There is no point in having all of this value chain of real-time insights and autonomy if you need to wait hours for the next satellite,” Lasserre emphasized. “To me, that’s the tipping point, where the value is just suddenly a lot higher commercially and for governments.”
