Voerrobot + AI: Closing the Loop From Feed to Yield

A feeding robot (voerrobot) automates the execution of feeding: it mixes rations, feeds small portions many times a day and keeps feed pushed up to the fence. What it cannot do is tell you whether that ration is actually working. AI closes that loop by connecting the robot's feed data with per-cow milk yield, health signals and feed costs — so the feeding strategy improves over time instead of just repeating itself.

What a feeding robot automates today

Automated feeding systems in the class of the Lely Vector, Trioliet's automated feeders and comparable machines from other builders have largely solved the execution side of feeding. They weigh and mix a ration per animal group, deliver fresh portions many times a day, monitor feed height at the fence and push feed so it stays within reach. The mechanisms behind the benefits are well documented in vendor material and dairy research: more frequent, fresher feed typically supports steadier intake and a more stable rumen, and the labour saving on a daily chore is obvious to anyone who has spent winters on a mixer wagon.

To be clear about where we stand: Crux Digits does not build robots and never will. The hardware vendors and machine builders own that ground, and Wageningen leads the research behind it. Our work sits in the layer those parties leave open on farms and agtech fleets: the data that all these systems produce but rarely share with each other.

The loop the robot leaves open

A voerrobot executes decisions; it does not evaluate them. The ration itself is still set by you and your feed advisor, and the evidence for whether it works arrives slowly and in pieces: milk yield per cow sits in the milking robot or parlour software, milk recording (MPR) results come in periodically, health events live in your farm management system, activity and rumination data in yet another app, and feed prices in your accountant's spreadsheet. In practice the feedback loop closes at the pace of a periodic advisor visit — with judgement and experience doing the integration that the software does not.

That is the structural gap: the feeding system knows exactly what was fed, the milking system knows exactly what came out, and typically no system connects the two at the level where decisions are made.

What AI adds on top of a voerrobot

Linking feed to results. The foundation is unglamorous: one combined dataset in which loaded rations, actual intake, refusals, per-cow yield, components, health events and weather sit on a shared timeline. Once that exists, a model can estimate how ration changes relate to yield and feed efficiency (voerefficiëntie) per group — with confounders such as lactation stage, season and heat stress accounted for rather than argued about.

Ration scenario support. Not a black box that replaces your nutritionist, but a tool that answers questions like: what happened to feed efficiency the last three times we changed the maize share? Which groups respond, which do not? The advisor stays in charge; the model supplies farm-specific evidence instead of sector averages.

Feed cost forecasting. Feed is typically the largest single cost on a dairy farm. Combining your own consumption patterns with price developments supports purchasing and contract decisions with numbers from your own operation rather than rules of thumb.

Early deviation alerts. Intake drops often precede visible health problems. A model that knows the normal pattern per group — corrected for weather and ration changes — can flag deviations days before they would stand out in a walk through the barn. The same logic applies to the machines themselves: robot logs carry early signals of wear that a predictive maintenance model can pick up before a breakdown at 05:30 does.

The data you already have

Most of the raw material is already being generated on the farm: loaded versus planned rations and feed heights from the feeding system, yields and components from the milking system, MPR results, activity and rumination from sensors, and public weather data. The practical hurdle is access — export options and APIs differ per vendor and per system generation, and some data only leaves the system as PDF or manual export. Mapping what is accessible, in what format and at what frequency is genuine work, and it is where an integration project starts.

Because we are vendor-independent, the result is a data layer that works across brands — and as with all our custom software, the client owns the code and the models. Your farm data trains a model for your farm; it does not disappear into someone else's platform.

A practical route: start small

The route that works is deliberately unspectacular. First, an inventory of what data each system can actually deliver — this is where most ambitions quietly die, so do it first. Second, combine two or three sources around one concrete question: for example, what does feed efficiency per group actually look like week by week, and what moved it? Third, a small proof of concept that answers that question with your own data before any talk of platforms. We run this as a fixed-price sequence — a €2,500 audit, a €20,000 PoC, production from €50,000 — so the decision to continue is made on evidence, not on a subscription already signed.

Honest limits

Some things this does not do. Effects on milk yield are confounded — lactation stage, weather, forage quality between silage cuts — so a model needs months of data before its estimates deserve trust, and it will show correlations that require an advisor's judgement to interpret. On a small herd, the gains from sharper feeding may not carry a custom project; the arithmetic is different for larger farms, multi-site operations and agtech companies building on fleet data. And none of it replaces the feed advisor or the farmer's eye. It replaces the folder of PDFs between them.

Frequently asked questions