Landscape

Where our work sits, and what it builds on. We would rather a reader see the neighbouring work and our place among it than take our framing on trust.

Orbital compute

A wave of companies and programs is putting data-centre hardware in orbit for continuous solar power and radiative cooling — among them Starcloud, Google’s Project Suncatcher  (tightly clustered satellites with optical inter-satellite links), Aetherflux, Sophia Space  and Axiom Space, whose first dedicated orbital data-centre nodes launched in January 2026. Most of the public discussion is about launch cost, thermal design and power. Our work is on the layer above: once the hardware is up, how does a fleet decide what runs where when its nodes eclipse, its downlinks come and go, and its view of itself is always somewhat stale?

Two things we measure follow directly from that framing. First, in a tight formation every satellite can see every other, so the hard constraints are energy, heat and ground contact, not network delay. Second, in a spread constellation the network is store-and-forward, and the dominant delay is waiting for a contact window — the regime where our delay-coordination results apply. See orbital compute constellations.

Scheduling compute on satellites

Energy-aware task scheduling for satellite edge computing is an active research area. Representative recent work includes sunlight-aware scheduling that decides whether to process in orbit at all (PHOENIX, arXiv 2407.07337 ), energy-aware scheduling for federated learning across constellations (arXiv 2409.14832 ), and decentralized, hardware-aware schedulers evaluated against measured edge-processor power (Equinox, arXiv 2604.19958 ). Our contribution is a benchmark environment where any scheduler — heuristic, market, central optimizer or learned policy — is scored on the same frozen sky, and a measurement of when a decentralized mechanism actually earns its complexity rather than an assumption that it does.

Distributed autonomy in flight

NASA’s Starling  mission flew four small spacecraft through demonstrations of distributed spacecraft autonomy, an ad-hoc crosslink network and autonomous formation flying, including the coordinated selection of observation targets across the swarm and file transfer across hour-long network interruptions. That is the flight evidence that decentralized coordination is feasible on spacecraft. Our work is complementary: simulation at fleet sizes and in regimes no flight demonstration can sweep, with every result scoped to the fidelity of the model behind it.

Spacecraft simulators and learning environments

High-fidelity spacecraft simulation is well served by established tools — NASA’s General Mission Analysis Tool (GMAT), and the open-source Basilisk  framework from the University of Colorado Boulder, whose BSK-RL  package provides reinforcement-learning environments for spacecraft tasking. Orrery is deliberately lighter: an analytic orbit and eclipse model validated against external references, built so that a contact plan for a constellation and a coordination experiment on top of it can be generated thousands of times on a laptop. Where higher fidelity matters, the right move is to cross-check against these tools rather than compete with them.

Delay-tolerant networking

Routing across scheduled, intermittent links is the domain of delay-tolerant networking and contact graph routing , developed for deep-space communication and increasingly applied to large constellations. Our delay oracle computes earliest-arrival times over a contact plan in the same spirit. Where our questions differ is in what is being carried: not bulk data, but the shared state a fleet coordinates on — and what goes wrong when replicas of that state diverge.

The theory of delayed coordination

That communication delay can destabilize consensus and flocking is classical: for linear consensus under a uniform delay there is an exact stability bound in terms of the network’s Laplacian (Olfati-Saber & Murray, IEEE Transactions on Automatic Control, 2004). Our delay papers measure the empirical counterpart at scale — nonlinear primitives, moving goals, stochastic loss and thousands of agents — and the natural next check — next on our list — is to set the measured collapse against that analytic bound.

Learned policies and world models

Much of the current energy in multi-agent coordination is in learned policies and in world models used to train them. Our benchmarks are built to host those as first-class submissions: the swarm suite already runs multi-agent RL comparators, and the constellation tasks are being exposed through standard multi-agent environment interfaces so that a learned policy and a hand-designed one are scored on the same frozen environment.