Four AMR factories in one week: what changed in a year

The visit format is always the same: assembly floor, burn-in area, parts warehouse, and a talk with engineers without the sales department in the room. We care about defect statistics, test protocols and how the factory handles warranty claims — not demo stands.

A year ago “pre-shipment run” meant switch on, drive, switch off. Now category leaders give every machine a 72-hour burn-in with telemetry logging, and internal line defect rates at two of the four plants are below 0.8% — last visit, honest numbers started at two percent.

Why this matters to a buyer and not just an engineer: a line's internal defect rate is the best predictor of how often you will call service in year one. A machine whose illnesses were caught on the bench arrives at your warehouse boring — and a boring machine is the goal in this business. Which is why the first document we request from a new factory is not the price list but the defect statistics for the last two quarters.

Localized lidars and controllers cut sensor costs by roughly a fifth — yet AMR price lists barely moved. The difference went into R&D and service infrastructure: factories stopped competing on price and started competing on cost of ownership.

Since we are standing in a workshop that builds both, let's settle the eternal terminology dispute. An AGV (automated guided vehicle) follows a fixed guide: QR codes on the floor, magnetic tape or reflectors. An AMR (autonomous mobile robot) maps the building itself and plans routes in real time, driving around people and pallets. These are not “old and new technology” — they are two tools for two kinds of warehouse.

A practical rule from our projects: if the routes have not changed in two years and will not change for two more — the AGV is honestly cheaper. If the warehouse lives in waves, zoning floats and people work the aisles — the AMR earns back its premium during the first seasonal peak.

Half the cost of a failed AMR project is not the machines — it is the integration that was first discussed after delivery. A fleet with no link to the WMS (warehouse management system) can do exactly one thing: drive around beautifully. For it to move orders, the WMS task queue must turn into fleet missions — and here is what to find out about the machine before signing.

• ▸An open fleet API — REST/WebSocket with documentation shown to you before the contract, not “after the NDA, maybe”.
• ▸Task format — who slices an order wave into missions: the WMS, a middleware orchestrator, or the machine firmware. The answer decides whose programmer the project needs.
• ▸The connector — who writes and maintains it: vendor, integrator or your IT. In writing, with a budget and a deadline.
• ▸A sandbox — a test loop where the integration runs before the machines arrive: a fleet emulator saves weeks on site.
• ▸Connectivity degradation — what the fleet does when Wi-Fi drops: finishes current tasks, parks, waits. “Everything stops” is not an acceptable answer for a freezer warehouse.

In our deliveries the integration scenarios are part of FAT: before shipment the machine must accept a task from a test WMS loop, execute it and report back correctly. That is cheaper than discovering incompatibility on an assembled warehouse floor.

How we test the software at the factory

Beyond the hardware, our inspection runs three software scenarios. First, degradation: we switch off an access point and watch what the fleet does after one minute, after five, and when the network returns. Second, updates: we ask them to flash one machine and demonstrate the rollback path. Third, logs: we request last month's telemetry export for any serial machine off the line. A factory that gets through these three in half an hour can go into a project; a factory where they trigger a meeting — not yet.

Hardware across the category has levelled out: on paper, the gap between a good and the best AMR is single percentage points. What decides the project now is the infrastructure around the machine — WMS integration, regional spare parts, incident response speed. Choose on that; everyone can do specs now.