Earth observation satellites run on rules that nobody wrote down cleanly. Power budgets, thermal limits, separation windows between shots — these constraints often live buried in engineering documents or high-fidelity simulators, not in the tidy mathematical models that scheduling algorithms expect. A new paper from arXiv tackles that gap head-on, treating feasibility itself as something to be learned in real time.
What's new
Researchers introduce Conservative Constraint Acquisition (CCA), a domain-specific procedure that learns which scheduling constraints are actually in play by querying a binary oracle — essentially asking "is this schedule valid?" rather than being handed the rulebook upfront. CCA is embedded in a Learn&Optimize (L&O) framework that alternates between optimizing under whatever constraints have been learned so far and issuing targeted queries to fill in the gaps. The conservative design philosophy matters: it avoids over-constraining the learned model, so the scheduler doesn't artificially box itself out of valid solutions.
Why it matters
The numbers are meaningful. On synthetic instances with up to 30 tasks, L&O cuts the average optimality gap from 65–68% (a priority greedy baseline) down to 17.7–35.8%. At 50 tasks, it matches or beats a two-phase acquire-then-solve baseline (FAO) — 17.9% gap versus 20.3% — while using 21 oracle queries instead of 100 and running roughly five times faster. For any real-world deployment where each oracle query means spinning up an expensive simulator or pinging a ground system, that efficiency delta is the whole game.
What to watch
The current model is deliberately simplified — pairwise separation and global capacity constraints only. Scaling CCA to the full constraint complexity of operational EO satellites, or applying the framework to other domains where constraints are similarly opaque (robotics, supply chain, air traffic), is the obvious next frontier. The constraint acquisition problem isn't new, but framing it as an active, interleaved part of the optimization loop rather than a preprocessing step is a cleaner architecture for real-world deployment.