Multi-aircraft atmospheric sampling.

The mission

A boundary-layer study, a plume transect, a gradient experiment. Four aircraft launch from one field site: two holding altitude steps over the ground station, one running the long crosswind leg, one chasing the plume wherever the morning takes it. The value of the campaign is precisely that the measurements are simultaneous — the same minute of atmosphere sampled at four points. Every platform streams instrument telemetry home in real time, because the scientists on the ground are steering the experiment leg by leg, and every platform must remain commandable at all times. The formation’s shape is dictated by the atmosphere. The radios were not consulted.

What breaks

A conventional architecture gives each aircraft its own link to the ground station — a star. The star fails exactly when the science succeeds: the moment the formation spreads to the geometry the experiment wants, the far member is at the ragged end of its radio’s reach, and every bank and climb modulates what little margin remains. Antenna patterns are not spheres; a turning aircraft shadows its own link with its airframe several times per orbit.

The deeper problem is that reactive failover is too late for a moving formation. By the time a link has measurably died, the aircraft that carried it has been flying away from recovery for seconds, samples have gone unrecorded at the ground, and the safety case has to explain a command-channel gap. Classical answers — more transmit power, taller masts, tighter formations — either fail against geometry or make the experiment worse to make the network better.

The architecture on this mission

Every platform and the ground station run the same node and form one encrypted mesh — no controller, no cloud, no configuration ceremony at the field site. Any aircraft can relay for any other, up to 8 hops, and relays cannot read the traffic they carry. The far member of the formation reaches home through whichever neighbour currently has the geometry, so the experiment’s spread stops being limited by the reach of any single radio: the formation itself is the transport fabric.

The load-bearing mechanism is prediction. Each node advertises its position and velocity to the mesh, and the scheduler projects the formation’s geometry forward — a link that is healthy now but attached to an aircraft accelerating away is already being scored down, and traffic migrates to the better route before the loss arrives, not after. Combined with links probed several times a second and reweighting within roughly half a second of measured change, a manoeuvring formation hands its routes over continuously and silently; the operators see instruments, not link alarms.

Traffic classes keep the order of importance intact when a route narrows to one thin relay hop: command and the sampling telemetry hold reserved floors, camera feeds and full-rate housekeeping shed first. The classification reads the traffic itself — autopilots and instruments fly unmodified.

How the sortie unfolds

1. Pre-flight. Four aircraft power up at the field site; the nodes authenticate by key and the mesh forms on the apron. One dashboard shows every link in the formation with its measured state — before anything flies.
2. Climb-out. The formation climbs in company. Every node is in direct reach of the ground station; every link is probed continuously and the mesh knows its own geometry from the shared position feeds.
3. The spread. The formation opens to the experiment’s geometry. As the crosswind aircraft approaches the edge of direct reach, prediction has already shifted its traffic through the mid-altitude neighbour — two short hops replacing one failing long one, with no dropout to trigger it.
4. The transect. The plume chaser departs downwind, banking through its own antenna shadows. Its routes migrate neighbour to neighbour as the geometry rolls; telemetry at the ground station stays continuous, and the command channel holds its reserved floor throughout.
5. The rejoin. The formation collapses back toward home; routes shorten to direct paths on their own, the same continuous re-scoring in reverse. Nobody has touched a radio all sortie.
6. The debrief. The per-link journal — round-trip, loss, route changes, timestamped — sits beside the science record. Every gap the instruments show can be answered from the link record, and the next sortie’s geometry is planned from measurement instead of superstition.

What each mechanism contributes

• Self-forming mesh with per-hop encryption — every platform a relay for every other, up to 8 hops, no central server; relays cannot read what they carry. Shipped.
• Position-aware predictive re-scoring — shared position and velocity project the formation’s geometry forward; traffic moves off a departing member before its link fails. Shipped, opt-in.
• Traffic classes — command and sampling telemetry reserved, bulk elastic, on whatever route is currently carrying. Shipped, opt-in.
• Continuous measurement — the entire sortie’s link behaviour recorded per hop, aligned to the science timeline. Shipped.

The honest boundary: Atlas cannot extend any radio’s physical reach — it composes the reaches that exist, and a formation that spreads beyond every pairwise link at once is out of contact until the geometry closes. Flight coordination, airspace, deconfliction and the instruments themselves are the institute’s.

What a pilot should prove

• Formation spread beyond single-radio reach with continuous telemetry from every member, carried over mesh relays.
• Prediction against reaction: measured packet loss during a scripted departure manoeuvre, with position-aware re-scoring on and off.
• Command-channel continuity for the plume chaser through a full transect, including its own banking shadows.
• A per-hop link journal aligned with the sampling record, good enough to plan the next campaign’s geometry from.

Four aircraft, one field site, two flying days. The evaluation format covers the structure.