Using AI and SDRs for Autonomous CubeSat Operations - Real-World Uses

yeah the “HAL 9000 in a shoebox” pitch always conveniently skips over the part where your EPS browns out because someone turned on the radio *and* the heater at the same time.

but to your question—there *are* a couple that sorta count. ESA’s Φ-sat-1 did onboard AI for cloud detection to avoid downlinking useless imagery, which is actually a sane use case. not running the spacecraft, just saving bandwidth. and some cubesats have played with adaptive modulation on SDRs, but it’s usually pretty conservative… more like “smart-ish radio” than anything resembling autonomy.

i think that’s the pattern: narrow, well-bounded ML tasks = works. anything that smells like “let the model manage ops” = nope, not when a single bit flip already ruins your day. curious if anyone’s seen SDR-based adaptive links that *actually* stayed stable over months, not just a demo pass or two?

yeah, “smart-ish radio” is about the ceiling right now. anything beyond that and you’re basically trusting a model that was trained on nice clean ground data to behave during a solar storm and a half-dead battery… what could go wrong.

Φ-sat-1 is the poster child because it *didn’t* try to be clever about spacecraft control. it just said “hey, clouds are boring” and saved downlink. same with SDR tricks—adaptive coding/modulation is fine until your classifier gets confused and suddenly your link budget math turns into vibes.

honestly feels like the real win isn’t autonomy, it’s just reducing dumb ops overhead. let the sat decide “this pass is trash, don’t bother transmitting” instead of burning watts yelling into noise. anything more ambitious still smells like conference slides more than flight heritage.