One detection architecture, from a camera lens to a farmer's phone
Camera → Camera Vision → AI Detection → Biocenza Platform → Farmer Dashboard. Every stage below reflects what ships at launch, nothing here is a hypothetical API.
The Camera Vision pipeline
Camera
Standard off the shelf cameras mounted along the growing rows, no proprietary hardware required.
Camera Vision
Continuous image capture across the greenhouse, feeding frames into the detection pipeline in real time.
AI Detection
On-device models identify crop, ripeness stage, and early signs of disease from each frame.
Biocenza Platform
Detections are aggregated per greenhouse alongside environmental and oxygen exchange readings.
Farmer Dashboard
Results reach the farmer as a clear read: what is ripening, what needs attention, and when to pick.
What the models actually detect
Standard off the shelf cameras feed continuous frames into on device models trained to read three things: crop stage, ripeness, and early visual signs of disease, the same read a good farmer makes by eye, at every row, all day.
Real time crop scanning
Replaces the daily manual walk through with continuous camera coverage of every row.
Early pre harvest detection
Tells a farmer what is ripening and when, ahead of the harvest window rather than at it.
Oxygen exchange monitoring
Tracks each plant's oxygen exchange for ongoing eco-monitoring of crop and greenhouse health.
Disease and stress detection
Surfaces early visual signs of disease so a farmer can treat a few plants instead of losing a row.
The prototype being built right now
A conveyor belt and AI guided robotic arm that harvest within a two to three day precision window, running on the same detection models that power the scanning platform. Two working prototype units are being built now.
| Component | Role in the prototype | Est. cost |
|---|---|---|
| Raspberry Pi 5, 8GB | Onboard controller running detection and arm control | $80 |
| Raspberry Pi AI HAT | Neural accelerator so ripeness detection runs on the unit itself | $70–130 |
| Raspberry Pi Camera Module 3 (×2) | Stereo vision for ripeness classification and depth | $70 |
| Six axis arm kit, bus servos and gripper | The picking mechanism | $260–500 |
| Servo driver & power distribution board | Drives the arm servos from the Pi | $25 |
| Conveyor section, motor and driver | Moves picked fruit clear of the arm | $150–300 |
| Aluminium frame & rail mounting | Mounts the unit over a tabletop growing row | $200–400 |
| Power supply, wiring, enclosure | Power and weatherproofing for greenhouse conditions | $80 |
| Spares, tooling, 3D-printed parts | Replacement servos and grippers during testing | $150 |
| Total build (two units + shared server) | $2,570–4,130 | |
Within a $7,000 total budget, with contingency remaining. Indicative retail prices for named components, not vendor quotes.
Honest limits
What these units will prove, and what they will not: they are being built to show the detection models can drive real picking hardware, and to generate the training data for the commercial system. They will not be field hardened, a deployment unit has to run a full season at speed, unattended and safely, with support and spare parts behind it, so the commercial version will cost more than the prototype.
Right sized, not scaled down
It will not cost anything close to the industrial benchmark, and it is not meant to. Harvesting robots priced at $120,000–250,000 are built for very large greenhouses in the Netherlands and the United States, where one machine covers hectares of high wire tomatoes. Korean strawberry farms are small, family run, and grown on tabletop rows, a different class of machine for a different customer, not a discount version of theirs.
One brain, many bodies
Intelligence stays on the server. The robots stay cheap. Every Biocenza microbot is a body without a brain: motors, a camera, a gripper and a radio. Detection, path planning and task scheduling all run on the same local server that already powers the scanning platform, so a farm adds capacity by adding another inexpensive machine, not another expensive computer.
01
Farm server
Detection models, task queue, path planning and fleet state. The only computer on the farm that thinks.
02
Local network
A private wireless link inside the greenhouse. No cloud round trip, so a command reaches a machine in milliseconds.
03
Microbot
Motors, camera, gripper, battery, radio. Executes instructions and reports telemetry. No onboard AI.
Cost lives in the brain, not the body
Industrial harvesting robots are expensive largely because each unit carries its own high end compute, sensors and controller. Moving that cost to one shared server means each additional robot is close to the price of its motors and frame.
One model, every machine
Because the server does the thinking, every robot on a farm runs the same detection models at the same version. Improving the model improves the whole fleet at once, with no firmware update per unit.
Repairable by the farm
A machine with no onboard intelligence is a machine a farmer can fix. Swap a servo, replace a gripper, bolt on a new frame. The part that is hard to replace never leaves the server room.
Fails safe, not silent
If a unit loses its link to the server it stops rather than improvising. Autonomy is deliberately not distributed to the machines.
Lifter
Moves harvested crates between growing rows and the packing area.
Up to 15 kg per trip
Porter
Carries tools, trays and input supplies along the row so a worker does not walk back.
Up to 8 kg per trip
Scout
Drives the row with a camera, feeding frames to the server when a fixed camera cannot see far enough.
Camera and sensor mast only
Harvest arm
The six axis picking unit on a conveyor, running the same server side detection models.
Single fruit picking
None of the microbot classes are deployed on a working farm today. The harvest arm and scout exist as prototypes being built now, and the lifter and porter are designed but not yet built. The fleet view inside the platform runs on simulated telemetry so the control surface can be designed and reviewed ahead of the hardware.
Commodity compute
Raspberry Pi 5 compute with an AI HAT accelerator and a local server, not industrial controllers or cloud dependency, keeping both hardware and running cost low.
Off the shelf cameras
No proprietary camera hardware required at launch, the same approach comparable monitoring platforms already use in commercial deployments.
A durable moat
Proprietary detection models and accumulating greenhouse sensor data compound over time. A formal patent and freedom to operate review is planned before commercial deployment.