Bridge and viaduct inspection.

The mission

A motorway viaduct strides a river valley on three tall piers. The principal inspection is due, and the faces that decide the structure’s verdict are the ones nobody can see from anywhere convenient: bearing shelves, girder webs, the soffit between beams, the pier caps under the deck. A drone flies them bay by bay, metres from the concrete, while the engineer follows the feed frame by frame — because the difference between a shadow and a crack is a decision worth millions, and the alternative to seeing it live is rope access at height over water.

What breaks

The deck is a lid. The same mass of steel and concrete that blocks satellite positioning under the span also blocks the radio path between the drone and the crew — and it does so precisely at the point of interest, deep between the girders, where the aircraft is flying closest to the structure with the least room for error. A conventional single-link operation meets this with the lost-link procedure: signal fades, the aircraft climbs out of the bay, and the inspection of that face is abandoned or nibbled at from angles that compromise the imagery. The sortie’s purpose and the sortie’s link fail in the same place.

The quiet cost compounds afterwards: what the engineer could not watch live gets reviewed from a memory card at the office, and the re-fly of bay 7 — the one where the image was ambiguous — needs a second mobilisation, traffic management and all.

The architecture on this mission

The fix is a second node placed where the geometry is winnable. A relay aircraft holds station off the span — clear of the structure, seeing under the deck from outside — or, where a second aircraft is not available, a portable node sits clamped to the deck edge or a pier cap. Either way, the inspecting drone’s short hop out from under the girders survives conditions the long path to the crew never could, and the relay carries the traffic the rest of the way. Every node authenticates by key, every leg is separately encrypted, and routes recompute continuously as both aircraft move — no switchover event, no reconfiguration mid-sortie.

Whenever both the relay path and the direct path exist — at the outer bays, near the abutments — the control class is duplicated across them, so the loss of either is a non-event; measured loss on two independent paths multiplies down, which is exactly what a link carrying manual flight metres from concrete should be spending its spare bandwidth on. Video sheds quality before continuity when a bay pinches the link. And in the deepest sections, where even the relay geometry gets marginal, full-resolution imagery queues store-carry-forward on the aircraft and drains the moment it clears the girders — the engineer’s live view may coarsen for a minute; the record of the face never has a gap.

How the sortie unfolds

1. Setup. Ground station on the bank, twenty minutes of checks. The relay aircraft launches first and takes station off bay three, level with the deck; the mesh forms as each node authenticates.
2. Outer bays. The drone works the first soffit faces with both paths alive — direct to the bank and via the relay — control duplicated across the two. The engineer follows frame by frame.
3. Deep in the span. Between the girders of bay four the direct path is gone. The relay hop carries everything; the live view sheds a little quality in the tightest pocket and full-resolution frames queue on board, draining each time the drone clears a beam.
4. Relay rotation. The relay aircraft’s battery runs down mid-span. Its replacement takes station before it leaves; routes shift to the new node without operator action, and the inspection does not pause.
5. The flagged face. Bay seven shows a diagonal shadow the engineer does not like. Because the feed was live, the re-fly happens in the same sortie, from two more angles — not in a second mobilisation next month.
6. The debrief artefact. The journal maps link state bay by bay: where direct held, where the relay carried alone, where imagery queued. The next inspection of this structure starts with a relay plan instead of a guess.

What each mechanism contributes

• Mesh relaying — the relay aircraft or deck-edge node is an authenticated member; two short hops replace the path the structure forbids, with per-leg encryption throughout. Shipped.
• Duplication of the control class — across relay and direct whenever both exist; independent-path loss multiplies down instead of adding up. Shipped.
• Ordered degradation — video sheds quality before continuity in the pinched bays; control never contends with the stream. Shipped, opt-in.
• Store-carry-forward — full-resolution imagery from the deepest sections is late by seconds, never missing; the structural record has no gaps. Shipped, opt-in.
• Continuous measurement — a bay-by-bay link record of the structure, produced free by the sortie, from which the next inspection plans its relay stations. Shipped.

The honest boundary: Atlas contributes the network, not the navigation. GNSS is still gone under the deck — position-holding between girders remains the aircraft’s problem and the pilot’s skill, and no communications layer changes that. The aircraft, the airspace and traffic-management approvals over a live motorway or railway, and the structural judgement on every frame are the operator’s.

What a pilot should prove

• Control continuity across defined under-deck bays where the direct link measurably fails, with the failure and the relay takeover visible in the journal.
• Imagery completeness for the full soffit, including the deepest bays — every planned face delivered at full resolution by end of sortie.
• A relay rotation mid-sortie with no operator action and no pause in the inspection.
• A bay-by-bay link record of the structure, good enough to plan the next inspection’s relay stations from.

One structure, two aircraft, one inspection day. The evaluation format covers the structure.