Subterranean · Underground Relay Chain
A tunnel is a pipe. So we send a beam down the pipe, and drop a node every eighty metres to do it again.
Carry a 100 Mbps map stream and a 20 bps heartbeat out of a place where radio does not work, using hardware cheap enough to abandon and small enough that a robot can carry thirty-two of them into a hole.
A CATACOMB ejects a node every 80 m of straight run and at every corner and junction. Each node self-rights, acquires its uphill neighbour in six to twenty seconds, and reports its link margin. The chain is the reason the map comes out while the robot is still going in.
It does two other jobs that matter more than they sound. Because every node is a static point that measures optical time-of-flight to its neighbours at ±3 cm, the relay chain is also the survey control network — it is what takes a 3 km CATACOMB traverse from 2 to 6 metres of absolute error down to 0.45 m RMS. And because each node keeps a geophone, a microphone, a magnetometer and a PIR running on 1.8 mW for fourteen months, the chain is a tripwire. If somebody walks past node 14 in March, you know.
"Radio does not work underground" is half true, and the half that is wrong leads people to the wrong architecture.
A tunnel is an oversized lossy dielectric waveguide. Below a cutoff set by the bore nothing propagates at all; above it, propagation is by attenuated modes. For a 3 m bore that means nothing useful below roughly 50 to 100 MHz, and efficient propagation only above about 300 MHz where the bore is many wavelengths across.
And then, at 900 MHz in a clean straight 3 m tunnel, the attenuation is about 0.05 dB per metre — four decibels over eighty metres. RF works fine in a straight tunnel.
The killer is not distance. It is corners. A single right-angle bend costs 20 to 40 dB. One corner eats the entire link budget, and tunnels are not straight. Add an RMS delay spread of 20 to 100 ns, which caps the coherent bandwidth at 2 to 10 MHz, and fast fading of 20 to 30 dB over half a wavelength — 17 cm at 900 MHz — so a moving node's link dies and revives every few centimetres.
Through the rock itself there is nothing to discuss. Skin depth at 100 MHz in rock at 0.01 S/m is 0.50 m. At 1 kHz it is 159 m — which is why mine-rescue through-the-earth systems live at a few kilohertz and get tens of bits per second through 250 m with a thirty-metre transmit loop. That is a heartbeat, not a relay.
So RF is a bearer, opportunistically, and never the bearer. An architecture that assumes otherwise has already failed at the first corner. There is one genuine consolation: the same physics that stops your signal getting out also stops it getting out. At 900 MHz your emissions are contained within about a metre of rock, so underground RF is nearly perfectly EMCON-safe against a surface collector. The threat is a receiver inside the tunnel with you.
| Cutoff, 3 m bore | Nothing useful below ~50–100 MHz |
|---|---|
| Straight run, 900 MHz | 0.02–0.10 dB/m → 4 dB over 80 m |
| One right-angle bend | −20 to −40 dB |
| RMS delay spread | 20–100 ns → 2–10 MHz coherent bandwidth |
| Fast fading | 20–30 dB over λ/2 = 0.17 m at 900 MHz |
| Skin depth, 100 MHz, σ = 0.01 S/m | 0.50 m |
| Skin depth, 1 kHz | 159 m |
| Through-the-earth magnetic induction | 1.5 kHz, 30–60 bps through 250 m, 30 m loop |
| Wall brine, mud, steel doors | Tens of dB, or simply opaque |
| Adaptive RF power when permitted | −10 to +5 dBm — the minimum that closes the link |
| Transmit optical power, 40 mW | +16.0 dBm |
|---|---|
| Transmit optics loss | −1.2 dB |
| Beam divergence, full angle | 8.0 mrad |
| Spot diameter at 80 m | 0.64 m |
| Receive aperture, 50 mm | 1.96 × 10⁻³ m² |
| Geometric coupling loss | −22.1 dB |
| Pointing and jitter, tracked | −3.0 dB |
| Receive optics and filter | −1.8 dB |
| Tunnel particulate, nominal dust | −4.0 dB |
| Received power | −16.1 dBm |
| InGaAs APD sensitivity, BER 10⁻⁹ | −36.0 dBm |
| Link margin | +19.9 dB |
What eats the margin
Dust. Light dust costs about 3 dB per hundred metres. Post-blast dust reaches 40 dB per hundred metres, which puts 32 dB into an 80 m hop and takes the link down with 12 dB to spare on the wrong side. Expect ten to sixty minutes of optical blackout after a breach. The chain responds automatically: drop to 2 Mbps, re-plan to 25 m spacing if it has spare nodes, then fall back to acoustic and queue in flash.
Corners. A single diffuse bounce off shotcrete at 1550 nm — reflectance about 0.12 — costs 28 dB and destroys collimation. Non-line-of-sight around one corner is usable to about twelve metres at 2 Mbps and no further.
Which is why a node goes at every corner and every junction, not only every 80 m.
That is the single most important deployment rule in the product. It is why a 32-node magazine does not cover 3 km, and it is a consequence of a reflectance figure rather than of anyone's opinion. We would rather size the magazine honestly here than have a robot run out of nodes at station 2,400.
| Coupling | Piezo stack coupled to the tunnel floor or wall |
|---|---|
| Band | 800 Hz – 3 kHz, BPSK, heavy forward error correction |
| Attenuation | α = πf/(Qv) → 0.20 dB/m at 2 kHz, Q = 60, v = 4,500 m/s |
| Rate | 20–40 bps at 200 m · 8 bps at 250 m · 2–3 bps at 600 m |
| Ceiling | Nothing useful beyond about 900 m |
| Coupling penalty | −20 dB on mud or a rubber-tyred haul road |
| Jam and DF resistance | Very high. Conventional SIGINT is not looking at 2 kHz in the rock. |
| Carries | Status, position, a found-it flag. Never a map. |
| Medium | 250 µm bare fibre, 3,000 m spool, 0.52 kg, 92 × 62 mm |
|---|---|
| Loss | 0.25 dB/km at 1550 nm → 0.75 dB over 3 km |
| Rate and latency | 10 Gbps, 15 µs at 3 km |
| Jam resistance | Total |
| Objection 1 | It gets cut — by rubble, by a vehicle, or deliberately |
| Objection 2 | It is a physical trail leading back to your entry point |
| Doctrine | Primary when the mission accepts the trail. Secondary when it does not. A commander's call, and the system supports both. |
| Mass | 340 g |
|---|---|
| Envelope | 96 mm base × 52 mm, weighted truncated cone |
| Self-righting | Passive, from any orientation on a hard floor, under 2 s |
| Battery | 2 × Li-SOCl₂ C cells, 62 Wh, 108 g |
| Listen-only draw | 1.8 mW → 14 months specified (3.9 y theoretical) |
| Full-rate relay | 2.6 W → 23.8 h continuous, or 30 days at 3% duty |
| Optical head | 2-axis MEMS steering, ±35° coarse / ±2° fine, 6–20 s acquisition |
| Store-and-forward | 8 GB flash, Bundle-Protocol-style delay-tolerant networking |
| Ranging | ±0.15 m single-shot · ±0.03 m after 4 s of averaging |
| Sensor payload | 3-axis geophone, microphone, magnetometer, PIR |
| Security | Keyed Revenant Mesh peer, revocable, zeroise on tamper |
| Ejection | Spring, 4 m/s, from the CATACOMB magazine; or hand-emplaced |
| Unit cost, today / target | $1,410 / $340. Not met. |
| Magazine | 32 nodes |
|---|---|
| Nominal reach | 2.4 km at 80 m spacing, less once corner nodes are counted |
| Per-hop latency | 0.4 ms processing + 0.27 µs propagation |
| 30-hop latency | 12 ms |
| End-to-end rate | 100 Mbps — a chain is not a shared medium, so rate does not divide by hop count |
| Acoustic fallback, end to end | 20 bps · about 8 s for a 40-byte message |
| Node loss, straight run | Skip to n±2 costs 10 dB. Survivable. |
| Node loss, at a corner | Fatal. No line of sight exists. |
| Post-mission | 14 months as a passive sensor field |
Thirty-two nodes into the CATACOMB magazine at the portal. Ninety seconds.
Node one at the portal, hard-wired to the surface segment of Revenant Mesh. It is the only node in the chain with an external tie. Everything after it is a peer with no master.
Eject at 80 m of straight run, and at every corner and junction. Each node self-rights, spiral-searches for its uphill neighbour, acquires in six to twenty seconds, and reports link margin so the operator sees the chain build itself.
Each node ranges to its neighbours, averages for four seconds, and publishes a ±3 cm baseline into CATACOMB's factor graph. In a long straight bore — the one geometry where a lidar has no information about how far it has moved — that range is the only thing standing between a survey-grade map and a useless one.
Map bundles flow out at line rate; tasking flows in. Thirty hops of latency is twelve milliseconds, which is imperceptible to an operator watching a facility assemble itself on a screen.
On a link failure the chain reduces rate, then re-plans spacing, then falls back to acoustic, then queues in flash. It does not stop. A robot 2 km in with three broken hops behind it is still delivering data — just later.
After extraction the chain stays as a sensor field for up to fourteen months. On command or on tamper the nodes zeroise. They are cheap enough to abandon and keyed so that a captured node reveals nothing about traffic it did not terminate.
A collimated beam down a pipe has almost no loss beyond divergence and scattering. That is not a clever trick, it is the geometry of the problem: the adversary dug you a waveguide and lined it.
Eye-safe 1550 nm at 40 mW closes an 80 m hop with 19.9 dB of margin through nominal dust. The whole rest of the design — the gimbal, the acquisition scan, the corner-node rule — exists to keep that geometry valid while the world moves around it.
This matters more than the link rate. The chain does not need to be end-to-end connected at any instant. Map bundles queue at a break and flush when the link comes back, in the same way an interplanetary network handles a planet getting in the way.
Designing for intermittent connectivity rather than for a circuit is the whole reason this works in a medium that breaks circuits for a living.
Anyone can drop repeaters. A node that is a fixed point and measures optical time of flight to its neighbours at ±3 cm is something else: it is a survey monument.
A range measurement to a node behind the robot along the tunnel axis directly observes the one direction a lidar cannot see. That is why BARROW and CATACOMB were designed together, and why a bought-in repeater bolted to a bought-in robot does not get you a 0.45 m map.
The entire concept depends on nobody caring about a node once it is emplaced. That sets a cost target, and the cost target is the hardest requirement in the product.
Current unit cost is $1,410 against a $340 target. A two-axis gimbal, a 1550 nm laser and an SoC do not reach $340 today. This is on the product page because it is the number that decides whether the concept is real, and hiding it would be dishonest.
Bring us a tunnel geometry and an EMCON state. We will bring the link budget, the corner count, and the unit cost we have not yet hit.
Every figure on this page is a design target derived from first-principles analysis, not a measured result. No BARROW node has been built. Where a number here is later measured and comes out worse, we will publish the measured number. No export-controlled information is published on this site.
