CropSCOUT:
three subsystems,
one field robot.

The MVP was a four-wheel rover the size of a suitcase. Farmers told us to make it bigger and let it carry more than one tool, so the V2 base is a tracked undercarriage designed from scratch in SolidWorks.

The assembly is fully defined: VTR056 rubber tracks on 100 mm idlers and drive sprockets, four split road rollers per side, machined wheel hubs and axles, a track carrier each side, a battery tray with restraints for the 24 V 20 Ah pack, tank cradle and hold-down bars, and a pump shelf for the sprayer module.

• Low centre of gravity: tank and battery sit inside the frame, not on top of it.
• Every attachment mounts to the same top plate — sprayer first, weeding module next.
• Motor nodes move from UART to CAN bus on a Teensy 4.1 controller.

CropSCOUT Vision is the ML / AI subsystem, running today as a web tool and next on the robot's onboard compute. Photograph one leaf or one fruit and it runs two stages:

• Crop & photo-type gate — is this a tomato, cucumber, apple or lemon, and is it a leaf or a fruit? A whole fruit is identified but never given a leaf disease.
• Condition screen — the closest covered condition for that crop, with a confidence score, alternatives, and a red AI-evidence overlay showing which regions moved the score.
• What to do next — field priority and first actions, written for the grower, not the lab.

Models are exported to ONNX and run in the browser with ONNX Runtime Web, so photos never leave the phone. We judge every model change against our own field photographs, not the validation split — that is where the studio-heavy public datasets fall over.

This is experimental screening, not a diagnosis or a pesticide recommendation. The tool asks for the team password while the model is in evaluation.

Spraying was the second-largest time cost and the single largest cash cost our farmers named. The sprayer module is what CropSCOUT does first with the base and the vision stack.

• 5 US-gallon tank in a cradle with hold-down bars, filled and rinsed through the top cap.
• 24 V diaphragm pump on its own shelf, fed from the same pack as the drivetrain.
• Solenoid valves and nozzle boom switched per nozzle, so a plant flagged by Vision gets dosed and its healthy neighbour does not.
• Dose, pressure and litres-per-row logged to the farmer's phone with the rest of the telemetry.

Once the sprayer is proven in the field, the next module on the same top plate is automated weeding — the job 14 of 30 farmers named as their single largest time cost.

Every PCB in the MVP was designed and assembled by us — schematic, layout, fabrication, bring-up and debugging.

The board pictured is the motor-driver node: an ESP32-S3 driving 4× TLE9201SG H-bridges over SPI, taking 5–40 V in and regulating to 3.3 V and 5.5 V on board. It is one of the 7 PCBs in the rover's distributed architecture; Asadbek led the wheel-motor subsystem. Firmware is MicroPython: MQTT over TLS for commands and telemetry, a UART daisy chain to the other boards, closed-loop control of the four motors.

Next revision: the main controller moves to a Teensy 4.1 — more headroom, CAN bus, real-time control — and a Jetson-class module joins the stack so Vision runs onboard.