Unit of decision
1plant
Every action is chosen for one plant, not for the whole field.
Cropite runs vision on Jetson units mounted on sprayers, weeders and drones, and decides what each plant needs before the boom passes it.
Pre-launch · no account needed to try the detector below
ExampleA toy simulation in your browser, not Cropite's model. A blanket pass would spray 100 % of the strip.
Unit of decision
1plant
Every action is chosen for one plant, not for the whole field.
Design target
<50ms
From camera frame to nozzle command, inside implement travel time.
Products
5
Perceive, act, measure, learn and orchestrate, on one shared model registry.
Draft plan
2027
First design-partner season: detect-and-spray on one crop.
The problem
The gap is not knowing what is wrong. It is acting on it, plant by plant, while the machine is moving.
Most spraying is broadcast: the whole field gets the same product at the same rate, because nobody can see each plant at boom speed.
Scouting imagery arrives hours or days after the pass, at field level. By then the weeds are bigger, the disease has spread, and the next pass is blanket again.
Meanwhile labour is short, inputs are expensive and the rules on chemical use keep getting tighter.
How broadcast spraying works
100%
of the ground gets product in a blanket pass, whether a weed is there or not.
Too late
Satellite and drone maps are processed after the pass, so they can only plan the next one.
Too uniform
A healthy plant and a weed in the same row get the same treatment.
Too few hands
Hand weeding and selective harvest depend on crews that are harder to hire every season.
The loop
Each product does one job. Together they close the loop from a camera frame to a measured result.
Folira · Perceive
Detection and segmentation on Jetson: crop, weed, pest damage, disease, nutrient stress and ripeness, from cameras on the machine.
ExampleValues are illustrative.
Targis · Act
Opens one nozzle for one weed, moves one weeding tool, picks one ripe fruit, within implement latency and inside an approved plan.
ExampleValues are illustrative.
Acreon · Measure
A farm twin plant by plant, and a ledger that links every action to its outcome on the next pass.
ExampleValues are illustrative.
Synora · Learn
Synthetic, labelled scenes for new pests, new crops and bad light, kept only if they improve results on real images.
ExampleValues are illustrative.
Agrona · Orchestrate
Agents that plan scouting and draft prescriptions from the twin; regulated steps wait for an agronomist.
ExampleValues are illustrative.
How a pass works
Steps 2 and 3 happen on the machine, with no network needed. Step 4 syncs when the machine is back in coverage.
1
A Jetson unit and cameras go on the boom or toolbar. Calibration maps each camera to its nozzles or tools.
2
Folira finds every plant in each frame and labels it: crop, weed, stress, disease.
3
Targis picks an action per plant and fires the right nozzle before the boom passes it.
4
Acreon keeps the action, the location and, on the next pass, the result.
Who it is for
Growers
Row crops, vegetables, orchards and vineyards that want to cut inputs without losing control of the pass.
Service providers
Contractors who cover many farms and need a spray record per field and per client.
Equipment makers
Makers of sprayers, weeders and robots who want per-plant perception and action inside their own machines.
Under the hood
| Layer | What runs there | NVIDIA technology |
|---|---|---|
| On the machine | Per-plant detection, targeting, nozzle control | Jetson Orin or IGX, DeepStream, TensorRT |
| Simulation | Implement commissioning, policy training | Isaac Sim |
| Training | Per-crop fine-tuning, synthetic data | TAO, Omniverse Replicator, DGX |
| Twin and analytics | Per-plant field twin, outcome ledger | Omniverse, RAPIDS |
| Agents | Scouting plans, prescriptions, approvals | NeMo, NIM |
Why the edge
At 15 km/h a boom moves about 4 metres a second. A round trip to the cloud is too slow for one plant, so the model runs on the machine.
For developers
{
"pass": "2027-05-14T09:42:07Z",
"row": 112, "plant": 457,
"class": "weed.amaranthus",
"confidence": 0.94,
"box_cm": [12.0, 3.5, 4.0, 4.5],
"action": {"type": "spray", "nozzle": 14, "ms": 40}
}Detection API
Folira emits a structured event for each plant; Targis adds the action; Acreon adds the outcome later. Export as JSON or as an as-applied map.
The record
ExamplePlant IDs, classes and results are illustrative.
Stewardship
Every nozzle opening is stored with its location, product and reason. Growers can export it for their own records and audits.
Questions
An agronomist or an engineer reads every message, not a sales queue.
No. Cropite is software plus an edge unit and cameras that mount on existing machines, or is built in by an equipment maker.
The first planned model pack is one row crop with common broadleaf and grass weeds. More crops follow with design partners.
You choose the fallback per field: spray anyway, skip, or stop and alert. Uncertain plants are also sent for labelling so the next model is better.
Not yet. Cropite is pre-launch. Every figure on this site is an example or a design target until pilots report real results.
Next step
Share camera footage or images from a recent pass. We run them through Folira and send back a per-plant report, labelled as an early example.