Vinea, slide by slide.

Vinea turns the way greenhouse crews work into robot skills, and runs them on one platform inside the greenhouses growers already have.

In a 2019 survey of over 1,400 Dutch growers, 11% of glasshouse growers used a robot of any kind, mostly for spraying, and 2.6% planned to invest in robotics. In a 2025 WUR test, GRoW picked half the trusses staff marked ripe; 40% of ripe trusses failed its picking conditions, and in 1.5 hours it stopped twice and needed a person six times.

eternal.ag has run one robot at Van Noord Growers since September 2025 and planned three for summer 2026. Four Growers reports up to 50% labour savings at Westburg. Each robot does one job, and the crop has to fit the machine.

What these robots get wrong are judgement calls: which truss is ready, which leaf comes off, where the whitefly starts. Crews make those calls all day, from experience built up on the job. That judgement lives in the crew, and the crew is the part growers can no longer hire. When a crew leaves at the end of the season, its know-how leaves with it.

That judgement can now be taught. Robot foundation models learn a new task from recorded demonstrations, and Physical Intelligence alone raised $400 million in 2024 to build them. What these models lack is data from inside working greenhouses, and nobody records it there, because the people with the skill are out on the crop rows.

Vinea records crews on their normal shifts, and the work itself doesn't change. For picking, deleafing and rehanging, the crew uses a handheld tool with the same soft fin-ray jaws as the robot and a camera looking along them, so every grip and cut is recorded from the robot's point of view.

For scouting, eye-tracking glasses and a microphone record where an experienced scout looks and what they say they see, so the recording labels itself. Each job's recordings train a robot skill on top of an open robot foundation model.

I built a simulation where a robot harvests a whole greenhouse. Now it's becoming real: a robot arm learns from demonstrations I record, and a camera model tells ripe from unripe tomatoes live, on real fruit. Next come ripeness checks on the gripper and a test greenhouse where the arm picks cherry tomatoes. Today I teach the arm by hand; at scale, crews teach it as they work.

Once deployed, every robot records its own work: each pick, each miss and each time a person steps in, so the whole fleet keeps improving. Each new crop starts from what earlier crops taught it, because picking a pepper has a lot in common with picking a tomato. The more greenhouses Vinea runs in, the less new data each new crop and each new greenhouse needs.

Growers rent Vinea the way they would hire a crew: a pilot season, then a monthly fee per robot with service included, always priced below the labour it replaces. Pricing stays provisional until after the pilot.

At Degla AI, a YC-backed drone startup, I built and tuned drones, precision work of the kind Vinea needs. I made the harvest simulation and I'm building the real robot myself, while studying Electrical Engineering at TU/e and bartending on the side. I've interviewed tomato growers on what labour costs them and why they haven't bought a robot yet, and built Vinea around their answers.

Connect with me on LinkedIn, or write to me directly.