Constraint-Aware Execution Planning for Hybrid Space-Ground Compute Workloads
Abstract
Low Earth orbit satellites generate roughly two orders of magnitude more data than they can downlink per orbit. Closing that gap requires deciding, for every workload, what runs on the satellite and what runs on the ground — under power, thermal, communication, and compute constraints that all vary as functions of time.
This paper presents Constraint-Aware Execution (CAE), a planning system that takes a workload specification and an orbital state and produces a feasible execution plan: which steps execute on which compute node, when each step runs, what data crosses the space-ground boundary, and how the plan respects power, thermal, and orbital constraints. CAE operates in four sequential phases — orbital environment modeling via SGP4 propagation, compute placement, data transfer planning with error correction, and scheduling within orbital windows.
We deploy CAE as a production API operating on live satellite data and show that feasible plans for representative workloads are produced in under two seconds. The system treats onboard computation as the primary mechanism for closing the data-downlink gap, and the planner as the primary surface where physical constraints meet workload semantics.
This summary is provided for indexing. The canonical abstract is on arXiv.
Method — Four Phases
Production Deployment
CAE is deployed at docs.rotastellar.com/cae and powers the planning layer of the RotaStellar platform. The same API is exposed in the live tracker, which generates plans for real satellites against current orbital state.
Cite
@misc{mitra2026cae,
title = {Constraint-Aware Execution Planning for Hybrid Space-Ground Compute Workloads},
author = {Subhadip Mitra},
year = {2026},
eprint = {2605.04052},
archivePrefix = {arXiv},
primaryClass = {cs.DC},
url = {https://arxiv.org/abs/2605.04052}
}