Communication in Harsh Environments

Communication is the connective tissue of any distributed intelligence, and in hostile or unstructured environments it is the thing that fails first. Unlike a datacenter network, these domains — from the deep ocean to interplanetary space — are actively hostile to signal propagation. Radiation, attenuation over distance, moving obstructions, and unpredictable interference degrade, distort, or sever the links that make a collection of agents into a swarm.

This is the constraint Arboria’s research exists to characterize, so it is worth being precise about what it does. The usual framing treats degraded communication as a performance penalty: coordination gets worse in proportion to how bad the channel is. That framing is wrong, and the way in which it is wrong determines whether a system works.

Two case studies

Consider a swarm of autonomous submersibles mapping a hydrothermal vent field. Saline water attenuates radio almost completely, so acoustic communication is the default — and acoustics bring their own difficulties. Sound travels at roughly 1,500 m/s, so a kilometre of separation costs two-thirds of a second one way. Multipath reflections off the seabed and surface produce echoes indistinguishable from signal. The vents themselves are loud. The channel is not merely lossy; it is slow, which is a different and worse problem, because it means every agent’s picture of its neighbours is systematically out of date.

Now consider a swarm exploring an unmapped cave system. Rock is very nearly a perfect barrier to radio. Line of sight exists briefly and is broken by every turn. Agents must form relay chains, or physically carry data to one another — a robot as a data packet. Here the problem is partition: the swarm is repeatedly cut into fragments that cannot communicate at all, and must reconcile when they meet.

Latency and partition are the two failure modes, and most environments deliver both.

Design patterns

Multi-modal communication. Combine acoustic, radio, optical, and ultrasonic links, and switch on link budget. Each modality trades range against bandwidth against energy, and no single one covers the operating envelope.

Delay- and disruption-tolerant networking. DTN abandons the assumption of a contemporaneous end-to-end path. The bundle protocol stores data at each hop, transfers custody so that no bundle is held by only one node, and binds addresses late, when a route actually exists. It was designed for interplanetary links and it applies unchanged to a flooded mine.

Store–carry–forward relays. Mobile nodes ferry data between partitions opportunistically. Bandwidth here is a station wagon full of tapes, and it works.

Progressive fidelity. Send compact intent and summary first; send bulk data if and when the channel permits. What the swarm needs to coordinate is small. What it has gathered is large. Conflating the two makes coordination hostage to telemetry.

Local-first autonomy. Agents act on local goals within a bounded deviation from the global plan, rather than waiting for instruction. Under partition this is not a design choice; it is the only option.

Elastic topology. Dynamic relay chains, leader election, redundant paths — because the node you were relying on is going to disappear.

Encoding intent so it can converge

A swarm that partitions and reconverges must reconcile divergent state without a central arbiter, and without a merge that depends on the order updates arrived in.

Conflict-free replicated data types solve exactly this. Express each piece of shared state as a structure whose merge operation is associative, commutative, and idempotent, and any two replicas that have seen the same set of updates — in any order, with arbitrary duplication — hold the same value. Because such structures form a join-semilattice , their finite products do too, so composites inherit convergence for free. Mission intent, a stigmergic pheromone field, and a depot inventory can all be built from the same primitives, and the convergence argument is made once. This is the design in ICCD.

Practical discipline

Define a link budget per modality with automatic handover thresholds, and add hysteresis to those thresholds or the system will thrash between radios. Implement DTN queues with priority classes and expiry, distinguishing perishable state from archival data. Tune forward error correction and interleaving to the channel’s actual burst-error characteristics rather than to its average error rate. Where contacts are predictable — orbital passes, submersible surfacing windows — exploit the contact plan as a schedule rather than discovering the topology by probing it. Provide offline-safe behaviours and watchdogs for isolation lasting minutes to days. And log per-packet energy cost, because in most of these environments energy per delivered bit is the binding constraint, not bandwidth.

What to measure

Packet delivery ratio, by modality and hop count. End-to-end latency at the median and 95th percentile, reported separately for intent and for bulk. Link availability, contact-window utilization, and topology churn rate. Joules per delivered byte, separated for relay and endpoint. And age of information for critical state — not whether the message arrived, but how old the belief it produced now is.

That last metric is the one that matters, and it is the one most systems do not track.

Failure modes

Partition collapse follows from over-reliance on a single relay: rotate the role, provide multiple paths. Energy brownout hits relays first, because they carry everyone’s traffic: route with energy awareness and duty-cycle. Buffer bloat in unbounded DTN queues turns a delay problem into a memory-exhaustion problem: cap per priority class and define drop policy. Clock drift silently breaks time-to-live and cryptographic validity: plan for timekeeping without satellite navigation, with mutual synchronization beacons.

Example architectures

Underwater vehicles: an acoustic mesh for control, optical bursts for high throughput at docking, and surfaced radio for backhaul.

Cave explorers: breadcrumb radio relays dropped in place, with store–carry–forward robots bridging collapsed segments.

Cislunar operations: contact-plan DTN with orbiters as scheduled custodians, while the local swarm gossips intent among itself. See orbital compute constellations.

The thing this page exists to say

Everything above manages the transport problem. None of it addresses what delay does to coordination, and the two are not the same.

Our measured result is that coordination quality does not fall smoothly as delay grows. It collapses through a boundary near 10–20 steps of delay measured against the task timescale. Three properties make this consequential.

The collapse is primitive-independent. Gossip consensus, flocking, and CRDT-intent propagation over a contact plan all fail at the same place. The design patterns on this page do not move the boundary; they determine whether a message eventually arrives, not whether an agent acting on a stale message can still coordinate.

It is invariant in swarm size, across the range we tested. You do not scale out of it.

And it is recoverable by anticipation. An agent that extrapolates where its peers have gone since it last heard from them recovers most of the lost coordination quality — and the cheapest predictor wins. Constant-velocity extrapolation beat a Kalman filter in every condition we tested, because bounded actuation keeps trajectories nearly linear over the prediction horizon and the filter’s process priors add variance they never repay. That result is Anticipatory Coordination via Peer-State Prediction.

The design implication is unglamorous and, we think, correct. Before choosing a coordination protocol for a harsh environment, compute the ratio of your communication delay to your task timescale. If it sits on the wrong side of the boundary, no protocol on this page will save you, and the only remedies are to slow the task down, to shorten the delay, or to make each agent model its neighbours’ futures.