Ports and yards are radio canyons

A container terminal is a structure built out of cargo. Steel boxes stacked several high and several deep form walls and corridors across acres of paved yard, and those walls are as opaque to radio as anything a civil engineer would put up on purpose. Metal reflects, blocks, and turns an open site into a set of canyons. A machine working the lane between two full stacks is inside one of them, and the canyon has no windows.

Terminals know this and mitigate it the ordinary way: survey the coverage, find the dead zones, add radios, plan around what is left. It works for a while. Then the ship works.

The obstacles in this environment are the cargo. They are moved, restacked, loaded, and discharged as the entire point of the facility, which means the radio geometry is not a property of the site but a property of the site's current state. A coverage map is a photograph of a landscape that is rearranged for a living. It is accurate the day it is made and decaying immediately afterward.

The coverage map has a shelf life

This is the difference between a hard radio environment and an unfixable one. A factory has dead zones in the same places every day, so a survey retains value and engineering around it is a finite job. A yard's dead zones migrate with the stacking plan. The survey is a description of a configuration that has already been dismantled.

Two consequences follow, and both are worse than a coverage survey suggests. The first is that faults become unreproducible. A machine that behaved strangely in one row cannot be tested in the conditions that caused it, because the stack responsible has sailed. Unreproducible faults do not get root-caused; they get attributed, usually to whatever is newest on the site. Any system that depends on connectivity in a yard will accumulate a reputation built out of incidents nobody can reconstruct.

The second is that coverage degrades exactly when the yard is busiest. A full yard has more steel in it, in more places, than an empty one. Peak operations and worst coverage arrive together, which is the opposite of the correlation an operator would design for, and it means the system's hardest moments and its least connected moments are the same moments.

Taken together, these say something specific about mitigation. Improving coverage in a yard is worth doing and will never be sufficient, because the thing degrading it is the facility performing its function. The only durable answer is machines that keep working when the link is not there, with connectivity treated as an optimization the fleet can exploit when present rather than a dependency it assumes.

A yard is one traffic problem

The second particular is that a terminal is not a set of independent machines that happen to share a paved surface. It is one traffic problem, worked by equipment that has nothing else in common.

Quay and yard cranes are enormous, slow to start, slow to stop, and constrained to fixed geometry. Yard tractors and shuttles are small, quick, and free-roaming. Inspection and gate systems are fixed installations that observe rather than move. Their sightlines, stopping behavior, footprints, and turn constraints differ so much that a planner cannot paper over the differences with one motion model, and a right-of-way convention that suits two of them will be wrong for the third.

Then there is the part no fleet controls at all. Road trucks arrive from outside the terminal, driven by people employed by someone else, running none of the terminal's software and obeying only the signage and the gate. Any shared picture of the yard has to include actors that will never be participants. That is a modelling constraint before it is an integration constraint: the world model cannot be a registry of cooperating agents, because a large share of the traffic never registers.

Vintage compounds it. Terminals accumulate equipment across decades and vendors, and none of it will be replaced to accommodate a new compute layer. Whatever coordinates this yard has to attach to what is here, tolerate machines that expose very little about their intent, and remain useful when a given class of equipment contributes nothing to the shared picture except its observed position.

Outdoors is a specification

Nearly everything in a terminal happens outside, and outside is a set of requirements rather than a setting. Salt air corrodes continuously, working on connectors, fasteners, and heat exchangers, so corrosion resistance becomes a design input rather than a maintenance activity. Wind loads matter for anything mounted high. Temperature swings drive condensation inside enclosures, which is a more common cause of electronics failure than heat itself.

Sensing degrades in ways that are specific and gradual. Salt film builds on a lens. Rain and fog change what a camera sees before they change what a person notices. Low sun across a wet yard produces glare that no exposure setting resolves. A perception stack here has to cope not only with hard conditions but with sensors that are slowly getting worse, and a machine that behaves confidently on a clean lens and unpredictably on a dirty one is a machine whose worst behavior is scheduled by the weather.

What the manifest carries

The data in a terminal has an unusual ownership structure, and it is the reason the export question is not the terminal's to answer alone. Most of what the systems know is not the operator's information. It describes shippers, consignees, and goods, held under contract, and it is simultaneously subject to customs and border regimes with their own handling rules. The terminal is a custodian. A custodian cannot unilaterally agree to send someone else's commercial records to a third party's infrastructure, whatever the terms of that infrastructure say.

Imagery has the same character in a less obvious form. Container photography is operationally necessary — damage assessment, seal verification, number reading, dispute evidence — and it is also a continuous record of who is shipping what, in what volume, on what cadence, through which berth. Aggregated over a season, that is a picture of trade flows that a shipper's competitor would value and that the shipper never agreed to publish. Video of a working yard is commercially sensitive in the same way process video is sensitive in a plant, except that the sensitivity belongs to customers rather than to the site.

Where the compute goes

A terminal raises a siting question that indoor archetypes do not. If the compute serving a fleet belongs near the fleet, and the fleet works on pavement, where exactly does it go?

There is usually a building — operations, gate complex, maintenance shop — with conditioned space, real power, and existing network. Putting the plane there is the straightforward answer, and it is the right one when the yard's internal network can carry coordination traffic from the working face back to it with predictable timing. The alternative, distributing compute into outdoor cabinets nearer the work, buys shorter and more reliable paths and pays for them in thermal design, corrosion exposure, service access, and physical security for equipment sitting in a yard that is not a locked room.

Neither choice is free, and the thermal side is genuinely unfinished. Sealed, filtered enclosures are the point of this design, and sealing fights sustained accelerator load in a cabinet under direct sun. How to trade density against service life in a cabinet that lives outdoors on a coast is not something we have settled. What is clear is that the answer has to be one an operator can maintain with the crews it already has, because a terminal is not going to staff a datacenter to run a robot fleet.

Designing for the gap

The design intent for this archetype follows from the coverage argument rather than from a preference for local compute. Edge nodes have to hold full autonomy through a gap, not a degraded mode, because the gaps are frequent, unpredictable in location, and correlated with peak activity. A yard-side plane holds the shared picture, coordinates across machine classes that share no abstraction, and keeps the heavy inference on the terminal's own network. Cargo data, imagery, and the yard's operational record stay inside the terminal boundary, where the custodial obligations that govern them can actually be honored.

The measure of the design is what happens in the canyon. A fleet that thinks only where the radio reaches is not autonomous in a container terminal. It is autonomous in the parts of the terminal that are currently empty.