FRAVEBOT

Digital twin of the greenhouse · synthetic training data for crop and navigation AI

Scout works in your greenhouse sooner, thanks to its digital twin.

We build a photorealistic copy of your greenhouse: rails, pipes, crop, people, trolleys and lighting. The robot's models learn on it first, so a new greenhouse, a new variety or a rare disease does not mean months of collecting and labelling photos.

THE TWIN IN 57 SECONDS · Digital twins of a tomato and a pepper greenhouse: plant detail, daylight and LED, synthetic scenes for navigation, and models trained on renders only, tested on real video from the robot.
Ready from day onemodels trained before the robot arrives
No manual labellingevery object comes with its exact outline
Works under your LEDstrained for your light recipe
Rare cases coveredrare diseases and situations, built in the twin

Built from your greenhouse

Your crop, your light, your rails

The twin is built from photos and videos of your own greenhouse, so the models learn on what the robot will really see.

The same tomato scene in daylight and under two horticultural LED recipes
RENDER · The same scene in daylight and with two LED recipes, at the same light level on the crop.

Daylight or LED, same scene

Horticultural LEDs mix red, blue, white and far-red light, and the mix changes during the day. We render the crop under each recipe, calculated for the colour sensitivity of the robot’s own camera, so the models are adapted to your light before the robot arrives.

Render of the head of a tomato plant and a close-up of the hairy stem
RENDER · The head of a tomato plant and a close-up of the modelled stem hairs.

Down to the hairs on the stem

Plants grow up the string into trusses from flower to ripe fruit. Fruit colours, leaves and stem bark come from real annotated photos; even the hairs on the stem are there, because they change how the stem looks to the camera.

Render of the pepper greenhouse with Scout on the main path and peppers on double gutters
RENDER · A pepper greenhouse twin, peppers on double gutters.

Peppers too

Same code, a different crop: double gutters, two stems per plant and plants up to about 3.5 m. Rails are modelled at the gauge of your greenhouse. Pepper fruits come from a 3D scan of real peppers, and the Nezara stink bug can be placed in all its stages.

Does it work on real photos?

Trained on renders, tested in the greenhouse

We train on images from the twin and test on real photos from the robot, taken on other days. A model that has never seen a real photo already finds the concrete, the rails, the fruit and the stem. A smaller set of real photos added on top makes it better still.

Tomato fruit detections on real greenhouse photos by a model trained on synthetic images only; white outlines are manual annotations
REAL · Fruit detection on real photos by a model that never saw a real photo. White outlines: manual annotation.
A crouching worker and a cardboard box detected in a real video from the Čerstvě utrženo greenhouse
REAL · The robot’s navigation camera in a real greenhouse. The worker and the box are found by a model trained on renders only.
What the model finds on real photosTrained on renders onlyRenders + some real photos
Concrete of the main path0.870.95
Rail in front of the robot0.800.90
Empty row0.750.97
Tomato fruits0.870.94
Stem near the plant head0.780.95

Scores from 0 to 1: 1 means everything found and outlined as a person would.

Pepper leaf photographed with a phone and cut out by Capture FRAVEBOT Capture · capture.fravebot.com Real photos from your team’s phones The twin is one source of training data, Capture is the other. It extends our datasets to places Scout does not reach: photos your team takes with their phone are checked and labelled, then added to the datasets Scout’s models are trained on, so Scout keeps getting better. For your team it also works as an identification tool, with an answer in seconds. See how Capture works →

Navigation

Navigation, practised in the twin

In a real greenhouse the robot meets a bucket at the row end, a trolley, a worker kneeling by the plants. In the twin we prepare these as scenarios in any number and variant. Scout also drives in the twin as a physical model with its real weight, drive and rail rollers, so how it turns, leaves the path and rides the heating pipes is tested before it drives into your greenhouse.

Synthetic images from the robot’s navigation camera: buckets at the row ends, workers on the path, a worker kneeling at a row and a worker crouching at boxes
RENDER · Renders from the navigation camera, about 20 cm above the concrete: buckets, workers on the path, a worker kneeling at a row and one crouching at boxes.
SIMULATION · Physics only, nothing animated by hand: Scout turns off the concrete path and its rollers line it up with the row by themselves. Twin of the Čerstvě utrženo greenhouse.

Where it ends

  • Real photos still help. Every result above improves when some real, hand-annotated images are added.
  • Leaf stalks and support strings cannot yet be learned from renders alone; with real photos added they are still weak.
  • LED spectra are modelled from the manufacturers’ descriptions for now; spectrometer measurements will replace them.
  • The twin shortens calibration for a new crop or variety. It does not remove it.

What's next

  • The twin compares the robot’s daily photos with its own renders and adjusts itself to the current state of your crop.
  • Navigation models trained on the scenarios decide whether to drive around an object, wait, or stop and report it.
  • Rare diseases and pests modelled before they appear in your greenhouse.
Render of the tomato greenhouse twin
Journal · Nov 20, 2026 · Vratislav Beneš A digital twin of the greenhouse: from a few plants in Omniverse to a whole greenhouse in Blender How we built it, why it was impossible three years ago, and where real images still help. Read the full story →

Send us photos of your greenhouse

A few photos and a short video of your rows, main path and lighting are enough to start your twin.

Get your Scout