Structural inspection in the urban canyon.

The mission

A storm has been through the city centre, and three buildings need their facades surveyed before the cordons come down: loose cladding, cracked render, a parapet the ground team cannot see from the street. The survey drone works its way around each structure at close range — up one face, around the corner, down the next — while the surveyor at street level watches the feed and marks defects frame by frame. The flight geometry is dictated by the structure. The radio geometry is dictated against it: every corner the inspection requires is a corner the link must survive.

What breaks

The vendor’s own video link is line-of-sight radio, and between buildings line of sight is binary. On the open face it is excellent; the moment the aircraft turns past the facade it is gone — not degraded, gone, because concrete and glass do not negotiate. A conventional single-link architecture treats each shadow as a disconnection: the stream resets, the operator waits, and the pass is flown again. Urban cellular inverts the problem: coverage is dense, but a low-flying aircraft is exactly the client the network was not planned for, seeing many cells at once and being handed between them constantly — and on a single modem, each rough handover is its own stutter.

The classical answers each give something up. Flying higher restores line of sight and destroys the close-range imagery the survey exists to collect. Repositioning the ground station for every face turns a one-hour survey into an afternoon. And “just use LTE” trades one failure mode for another, with no fallback when the one modem lands on a congested or fussy cell.

The architecture on this mission

The aircraft’s node bonds both links it already has — the vendor’s stock video link and an LTE modem — into one encrypted tunnel to the ground station. The bond is vendor-neutral: the video link is just another measured path, probed several times a second like everything else, and the autopilot and camera are unmodified. When the aircraft is on the open face, the direct link carries the weight and the cellular quota rests; when it turns past the facade, the traffic is already flowing on cellular and the stream never resets.

“Already flowing” is the load-bearing part. Where the radio reports its signal level, Atlas projects the trend — a facade approach is a fade with a shape, and predictive re-scoring shifts traffic off the dying link before the first lost probe, not after the operator sees the freeze. On the cellular side, endpoint roaming absorbs the constant cell handovers and address changes of low-altitude urban flight: a modem that re-addresses mid-pass is a non-event inside the tunnel, not a reconnection. Both ends sit behind ordinary networks, joined through a relay node reachable from anywhere — nothing for the building owner or the carrier to provide.

Traffic is classed on top: the command channel holds a reserved floor on whichever paths currently exist, and the video stream sheds quality before continuity when a corner briefly leaves only a thin path. A four-megabit stream can never crowd out the few kilobits that steer the aircraft next to a structure.

How the survey unfolds

1. Setup at the cordon. The surveyor powers the aircraft at street level; the node bonds the vendor link and the LTE modem into one tunnel. One dashboard shows both links’ measured state before the first metre is flown.
2. The open face. Line of sight is clean and the vendor link carries the survey; the cellular path idles at a trickle, measured and ready, spending almost nothing of its quota.
3. The corner. The aircraft turns past the facade. The vendor radio’s signal level has been falling for seconds, and the projection moved the stream onto cellular before the link died — the surveyor sees the same feed, uninterrupted, and marks the next defect.
4. The courtyard. Deep between buildings, the modem is handed roughly between cells and re-addresses. Endpoint roaming carries the session across it; the pass continues instead of restarting.
5. Back into view. Rounding the last corner, the vendor link’s measurements recover and it re-scores its way back into the bond — no switchover event, no operator action, cellular spend dropping back to a trickle.
6. The debrief. The event journal holds a per-link, per-pass record: where each facade shadowed the direct link, how deep, and what the bond did about it. Flown once, that record is a radio map of the structure for every future survey.

What each mechanism contributes

• Vendor-neutral bonding — the stock video link and an LTE modem as one tunnel with one address; the aircraft and camera unmodified. Shipped.
• Predictive re-scoring — signal-level trends projected so traffic leaves a fading link before it fails, not after the freeze. Shipped.
• Endpoint roaming — cell handovers and address changes absorbed inside the tunnel; a re-addressing modem is not a reconnection. Shipped.
• Class floors and ordered degradation — command reserved, video elastic, on whatever the corner currently allows. Shipped, opt-in.
• Continuous measurement and the journal — a per-link record of every shadow and handover, per pass, for the debrief and the next survey. Shipped.

The honest boundary: Atlas cannot put radio where physics forbids both links at once — a deep courtyard that blocks the vendor link and has no usable cell is a genuine hole, and the answer there is a relay node at the corner or a second aircraft joining the mesh, which the architecture accommodates but the operator must field. The aircraft, the airspace approvals and the inspection judgement are yours.

What a pilot should prove

• Stream continuity through a counted set of corner transitions on a real structure — bonded versus each link alone, from the built-in measurement.
• Predictive behaviour made visible: the journal showing traffic leaving the vendor link ahead of measured failure on facade approaches.
• Cellular handovers absorbed without a stream reset across a full low-altitude survey.
• A per-pass link journal usable as a radio map of the structure for planning repeat surveys.

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