Planet Labs runs inference in orbit
- Planet Labs said on April 7 it ran AI object detection directly aboard its Pelican-4 Earth-imaging satellite, shifting analysis from ground stations into orbit. - The demo used an NVIDIA Jetson module to spot aircraft near Alice Springs, Australia, with about 80% initial accuracy on raw imagery. - That matters because faster onboard triage can cut bandwidth use and move Planet toward near-real-time geospatial alerts.
Earth-observation satellites usually work like flying cameras. They collect huge image files, dump them to the ground, and wait for servers on Earth to do the thinking. Planet just showed a different model. On April 7, 2026, the company said its Pelican-4 satellite ran an AI object-detection model in orbit and identified aircraft before the data ever came home. (businesswire.com) ### What actually changed in orbit? The important shift is simple — the satellite did inference onboard instead of acting as a passive sensor. Planet said Pelican-4 processed raw imagery in space using NVIDIA Jetson hardware and produced detections of airplanes near Alice Springs, Australia. That is the step people mean whe(businesswire.com). (businesswire.com) ### Why is that a big deal? Because the bottleneck in Earth observation is often not taking the picture. It is moving, sorting, and interpreting the data fast enough to matter. A satellite can see something urgent — a runway filling up, a ship appearing, flood damage spreading — but if the image has to be fully downlinked a(businesswire.com)ortant part first — basically the “something changed here” message instead of the whole haystack. (businesswire.com) ### What did Planet’s demo actually do? It was not a vague AI claim. Planet said the model detected airplanes on the ground from Pelican-4 imagery and generated outputs like geo-rectified files and object data. The initial accuracy was about 80% on raw images, which is good enough to prove the pipeline works but still clear(businesswire.com) on Earth. (businesswire.com) ### Why start with aircraft? Aircraft on a runway are a clean first target. They are discrete objects, strategically important, and useful for both civil and defense monitoring. If a satellite can reliably flag planes, the same architecture can later expand to ships, vehicles, infrastructure changes, wildfire signatures, or(businesswire.com)ecome selective about what it sees. (spectrum.ieee.org) ### How does this fit Planet’s bigger plan? Planet has been building toward a more real-time network for a while. Its Pelican line is the higher-resolution tasking fleet, and the company has also been talking about future Owl satellites with onboard AI as a core feature. Planet framed this milestone as part of a broader push toward “Planetary Intelligence” — t(spectrum.ieee.org)and flag events quickly. (businesswire.com) ### Is this really the first time anyone has done this? The careful version is: Planet called it one of the first times an Earth-imaging satellite has moved beyond simple capture to onboard AI inference and analysis. That wording matters. Spacecraft have done onboard processing before, but doing useful computer-vision infer(businesswire.com)s in space” and more “commercial imaging satellites are starting to think before they transmit.” (businesswire.com) ### What is the catch? Onboard inference saves time and bandwidth, but it also forces hard tradeoffs. Space hardware has tighter power, thermal, and compute limits than a ground data center. Models have to be smaller, tougher, and more carefully tuned. And if the detector misses something or flags the wrong thing, the satel(businesswire.com) making orbital triage dependable enough for customers to trust. (spectrum.ieee.org) ### Bottom line? Planet did not just speed up image processing. It tested a new division of labor between orbit and Earth. If that keeps working, satellites stop being cameras with delayed answers and start acting more like live sensors with judgment. (businesswire.com)
