Melkrobot Data: What Your Robot Knows About Every Cow

A milking robot (melkrobot) records dozens of measurements per cow at every milking: yield per quarter, milk conductivity, flow rate, milking duration and interval, refused visits, concentrate intake — and, via collar sensors, activity and rumination around the clock. With AI models trained on your own herd data, that same data supports early mastitis signals, sharper heat detection, yield forecasting and better culling decisions. Today, most of that value stays locked inside the vendor platform.

What your milking robot registers at every visit

Modern milking robots — Lely Astronaut, DeLaval VMS, GEA DairyRobot — are remarkable machines. To milk a cow safely and efficiently, they have to measure her continuously. A typical visit produces yield per quarter, electrical conductivity of the milk (geleidbaarheid), milk flow and milking duration, the time of the visit and the interval since the last one, refused visits, and the amount of concentrate (krachtvoer) dispensed. Many farms add collar or ear-tag sensors on top, contributing activity (activiteit) and rumination (herkauwactiviteit) day and night.

Multiply that by two to three milkings per cow per day, across a full herd, across seasons and lactations, and you are sitting on one of the richest operational datasets in Dutch agriculture. The robot vendors know this: their management platforms, such as Lely Horizon and DeLaval DelPro, use it to run the robots and generate attention lists. That is exactly what those platforms are designed to do, and they do it well.

Why the data mostly stays inside the vendor platform

The limitation is not the hardware — it is scope. A vendor platform is built to operate that vendor's equipment reliably on thousands of farms at once. Its alerts therefore typically rely on thresholds and rules tuned to work acceptably everywhere, which also means they are rarely tuned specifically to your herd, your barn and your feeding strategy. And the platform only sees its own world: the robot's measurements, not your feed system, not your herd management records, and not the quality results coming back from the dairy processor.

The result is a data silo. The robot knows things the feed computer does not; the herd management system knows things the robot does not. The interesting questions — why is this cow's yield drifting, which cows deserve extra attention this week, what will next month's tank volume look like — live across those systems, not inside any single one. That cross-system layer is what a vendor-independent data and AI partner for agriculture builds. It complements the robot; it does not replace it. We do not build robots — we build the intelligence layer on top of the data they already produce.

What AI on top of melkrobot data makes possible

Early mastitis signals. Elevated conductivity in one quarter, combined with a yield dip and changed milking behaviour, is a classic early pattern. A model trained on each cow's own baseline can flag deviations earlier and with fewer false alarms than a fixed threshold applied to the whole herd, because 'abnormal' is then defined per cow, per quarter, per lactation stage — not per population average.

Heat detection support (tochtdetectie). Activity spikes, changed visit behaviour and short-term yield dips together sharpen the timing of insemination. Collar systems already do this well on their own signal; combining sources typically reduces missed and false heats further — which feeds directly into calving interval and insemination costs.

Yield forecasting. Per-cow lactation curves, plus feed and seasonal data, support forecasting tank volume days to weeks ahead — useful for feed purchasing and planning. Just as important: a cow falling below her own expected curve is an early, unspecific health signal worth a look. The same anomaly-detection thinking that powers predictive maintenance in industry applies to a lactation curve — learn the normal pattern, flag the deviation early.

Culling and health decision support. Combining yield history, health events, fertility results and age into one transparent per-cow overview supports the hardest recurring decision on a dairy farm. Not a model that decides — a model that puts the evidence for each cow in one place, so the decision is made with the full picture.

The practical route: access, data quality, a small PoC

Step one is data access. Vendor platforms offer reporting and export functions, and increasingly APIs, to get your data out. You do not need to leave the vendor platform — you need a copy of your own data flowing to a place you control. Under typical arrangements the farm's data is the farm's; exercising that in practice is mostly a technical task, not a legal battle.

Step two is data quality. Sensor gaps, cows that swapped collars, health events that never got registered — every real herd dataset has them, and they decide whether a model can work. A short, fixed-price data audit answers whether your data can carry a model before anyone builds one. That is why we start every engagement with a €2,500 audit rather than a proposal full of assumptions.

Step three is a small proof of concept on your own herd data: one question — for example, early udder-health flagging — measured honestly against what the standard alerts already caught. At Crux Digits a PoC is a fixed €20,000, the client owns all code and IP, and everything is built GDPR-aware and with the EU AI Act in mind, as part of our custom software development practice in the Netherlands. If the PoC does not beat your current alerts, you know that too — on a fixed budget.

What the sensors do not replace

Honesty matters here. No conductivity sensor replaces the stockman's eye, and no model should make treatment decisions. What a good model does is triage: it turns thousands of measurements per day into a short attention list, so your experience is spent on the right cows at the right moment. False positives will happen, especially early on; a model earns its place on the farm by being right early a few times, not by being trusted blindly on day one. The robot collects the gold. Whether it stays ore or becomes value depends on what you build on top.

Frequently asked questions