IASTAM 6.0 · Track 1 · Problem 2
Onboard AI that measures its hardware, then adapts to the mission.
Without cost feedback
64missed deadlines
MIZAN, cost feedback on
?missed deadlines
The trade: 52 fewer missed deadlines for 0.6 points of expected accuracy (90.7 % to 90.1 %).
01 · The idea
A satellite that runs AI onboard must decide, frame by frame, how much of a small and shifting energy, memory and time budget each inference deserves.
MIZAN (ميزان, Arabic for balance) weighs accuracy against the energy and time it costs, using costs measured on the board it runs on, and keeps correcting them while it flies. Its safety filter can refuse a configuration but never chooses one.
Battery, temperature, deadline, memory and the value of the frame.
The most valuable option the energy price allows, from the corrected menu.
Refuse anything whose bound breaks the deadline, battery or heat limit. It can veto; it never chooses.
Run the chosen exit and resolution on the board, or skip the frame.
Time what the board really took.
Fold measured over predicted into a conformal latency bound for the next frame.
02 · Results
Measured Simulated
−31 %mean latency
After online cost correction, on the target board. The controller stopped trusting the laptop menu and spread its choices across seven options instead of one.
A cost menu profiled on a development laptop underestimates what the payload board takes, floor to floor.
Adaptive routing against full depth, at 64 px, with idle draw included, while the accelerator is saturated. Paced at 10 frames/s, the saving disappears.
| GPU | Full depth | Adaptive | Saved, saturated | At 10 fps |
|---|---|---|---|---|
| A40 | 324.4 mJ | 151.5 mJ | 53.3 % | not run |
| L4 | 157.2 mJ | 67.7 mJ | 56.9 % | 0.65 % |
| T4 | 157.8 mJ | 75.8 mJ | 52.0 % | −0.05 % |
03 · Demo
A four-minute narrated demo. A simulated 3U CubeSat has 40 ms per frame. Over a ground station the downlink slows inference 2.4×, then come heat and an eclipse with one solar wing stuck. The repository's own controller decides every frame, with and without cost feedback. Then the whole system runs live on the board, 1,440 frames per arm, and is killed mid-mission and recovers. Orbit sequences are simulations; board numbers are measured on a Raspberry Pi 4B.
| Controller | Missed | Mean latency | Exp. accuracy |
|---|---|---|---|
| MIZAN | 12 | 20.2 ms | 90.1 % |
| Menu-trusting | 64 | 22.0 ms | 90.7 % |
| Static, richest | 88 | 20.9 ms | 90.7 % |
| Static, cheapest | 0 | 12.9 ms | 86.4 % |
04 · Poster
05 · Paper
Ahmed Hedhili, Salem Essid, Ahmed Mnasri · IASTAM 6.0
We test the three levers an adaptive inference system has on a satellite: which model to run, how often to run it, and how much detail to process.
The model lever saves 52–57 % of energy per inference only on saturated accelerators. A laptop-profiled menu is 6.5× optimistic on the Raspberry Pi 4B, and correcting it online cuts missed deadlines by 49 % on the board. Live on the board, the full system held its latency bound. In a pre-registered held-out evaluation over laptop costs it caused no breach in 100 simulated runs, though at the board’s costs only with a battery reserve added afterwards, and it lost quality under heat; a step-down rule written afterwards beats every single fixed option in simulation (94.3 % against 92.3 %).
@misc{mizan,
title = {{MIZAN}: Measure-First Adaptive Inference for
Energy-Constrained Satellite Payloads},
authors = {Ahmed Hedhili, Salem Essid, Ahmed Mnasri},
note = {IASTAM 6.0 Technical Challenge, Track 1, Problem 2.
Team INTERSTELLAR, SUP'COM, University of Carthage},
year = {2026}
}
06 · Team INTERSTELLAR
École Supérieure des Communications de Tunis (SUP'COM), University of Carthage, Tunisia.
