Swarm intelligence for systems that cannot all talk at once
Arboria Labs studies how local rules and decentralized coordination produce robust collective behavior at scale — and, specifically, what happens to that coordination when agents cannot see each other in real time. Every swarm has the same problem in some form: links are limited, messages arrive late or not at all, energy is finite, and every decision is made on information that is already old.
We build the simulation stack, run the experiments, and publish what we measure. Every number we publish names the experiment that produced it.
Agent Coordination // Decentralized Awareness
Swarm intelligence and distributed systems are two halves of one problem: coordinate many agents under local information, bounded communication and partial failure. We work across both — flocking, consensus and coverage on one side; CRDT-based shared intent, delay-tolerant networking and decentralized markets on the other — with instrumentation that maps local parameter choices to the collective behavior they produce.
Where we apply it
Coordination under communication delay is the lab’s core science. Take away real-time visibility and decentralized coordination collapses through a narrow band of delay — and cheap prediction pushes that boundary back. Those results apply to any swarm whose agents talk over imperfect links: ground robots, drones or spacecraft.
Satellite constellations are where that science is being tested hardest right now. As compute moves to orbit for uninterrupted sunlight and radiative cooling, constellations have to schedule work, share intent and hold formation over links that open and close as satellites move. We model them with a constellation digital twin and test decentralized, energy-aware scheduling on it. See the console →
The Integrated Arboria Framework
Four engines, each doing one job:
- Leviathan — a C++ simulation core that steps agents under physics, a range-limited communication model with real bandwidth and energy costs, and fault injection.
- Gossamer — the coordination primitives, tasks, predictors and metrics: what each agent does, and how well the collective did.
- Orrery — the constellation twin. Orbits, inter-satellite visibility, eclipse and power, ground passes, and the delay oracle that says how long a message really takes to cross the constellation. It is validated against external references (the sun-synchronous inclination, the ISS’s nodal regression, a real ISS orbit from its published elements), not against itself.
- Maneuver.Map — runs the experiments, records their provenance, and replays them in the browser.
Together they power the Arboria benchmarks: versioned swarm and constellation tasks. Every result names the exact frozen environment it ran on, and a degenerate row or a fault that never fired fails the run instead of reading as robustness.
What we have measured
These are simulation results about coordination algorithms, scoped to the fidelity of the model behind them. They are not flight or hardware validation. See Reproducibility.
- Coordination has a delay cliff. Decentralized coordination quality collapses through a narrow band of communication delay — 10–20 steps in our studies — identically across gossip, flocking and intent-propagation primitives and at 500 or 2,000 agents. The phase diagram →
- Cheap prediction moves it. Extrapolating each peer’s last known state recovers much of the lost coordination past the cliff. Anticipatory coordination →
- Freshness matters; throughput does not. Cutting delivered messages fifty-fold barely moves coordination; making them older collapses it. Freshness vs throughput →
- In orbit, delay is waiting. On a Walker constellation with 5,000 km optical links, more than 99.7% of end-to-end delay is waiting for the next contact window rather than light travel — which puts constellations past the cliff above. Why →
- Markets are regime-conditional. On a terrestrial construction task an auction beats first-come-first-served only under scarcity — and loses when over-provisioned. We published that correction against our own earlier headline. The market paper →
Results on swarm and constellation scheduling, intent routing, formation safety and radiation are written up and will be published as they clear review. How this relates to Starcloud, Suncatcher, NASA Starling and open spacecraft simulators →
How the stack fits together
Our Research Ethics
Autonomous collectives have a specific set of risks that are not addressed by generic AI ethics. We take them seriously:
Dual-use caution. Swarm methods can be used for surveillance, denial, and coercion. We do not contribute directly to lethal autonomous weapons and we gate publications that could materially accelerate them.
Orbital and planetary responsibility. Large deployments in space create Kessler-syndrome risk and planetary contamination risk. Our simulations and recommendations treat debris, deorbit, and forward-contamination as first-order constraints.
Reproducibility over spectacle. Every published result names the batch it came from and records the environment that produced it. Our engines are proprietary, which limits what external reproduction can mean; we say so plainly rather than promising more than we can deliver. See Reproducibility and Data Availability.
Careful claims. Our results come from a simplified kinematic simulator. They are claims about coordination algorithms, not validation of physical devices, and we scope them to the fidelity that backs them — anchoring against known external results wherever we can.
Learn more about our principles →
Latest Updates
Releases, research notes, and corrections — including the ones that went against us.
Arboria Labs: swarm intelligence and distributed systems research, measurement-first.