Basil Bot

A small robot whose job is to find a basil plant, tell leaf from stem, and cut the leaf. Marin Marian’s master’s thesis in AI at Vrije Universiteit Amsterdam, supervised by Tony Currà.

The goal

Herb harvest is still mostly hands. Basil Bot was a test of whether a low-cost arm can do the narrow job: look at a plant, pick a leaf, and cut it. Not a greenhouse fleet. One plant, one cutter, on a bench at VU StartHub.

That is a different problem from Gibbon Bot (trees, not herbs) and from the survey drone (looking, not cutting). Here the camera has to drive the motors.

The team

• Marin Marian — master’s thesis in AI at Vrije Universiteit Amsterdam. Computer vision and visual servoing: tell stem from leaf in the camera image, then drive the joints toward the cut.
• Tony Currà — supervisor, hardware, and firmware. He designed and 3D-printed the parts, wired the servos, and made the arm move from a board you can power off USB.

The split on paper is hardware versus vision. In the room they worked the small problems together: a stem that moved, a servo that overshot, a cut that only works one way. Limited kit, a real job. That was the thesis.

The arm

Four axes — that is the main constraint. The scissors only close in a vertical cut. A leaf at an awkward angle, or a stem the tool cannot approach from above, is out of reach. A six-axis arm would shrug and rotate the wrist. This one cannot. So the interesting work was not buying another motor. It was which leaf to pick, how to approach, and how to aim a tool that only cuts straight down.

Hobby servo motors, chosen because they are cheap and they run on low power — the whole machine is USB-powered, camera included. No industrial drives, no 48 V bus. That keeps it on a desk. It also means backlash, slack, and not much torque. The firmware has to live with that.

Every structural part is printed. The transmission is printed gears, in the open, where you can see them mesh. The cutter is printed too — a small scissor at the wrist. That is the point of this machine as a prototype: you can watch how motion gets from a servo horn to a joint, instead of hiding it in a black gearbox.

The eye is a Seeed Studio XIAO ESP32S3 Sense — an ESP32 board with a small 2D RGB camera, about €20. Power and data go over the same USB cable. No stereo pair, no depth sensor, no global shutter. One noisy colour image, looking along the arm toward the plant.

Seeing and reaching

The camera is the hard part. A €20 module is grainy, it rolls, and it has no sense of distance. Stem and leaf are both green. Marin’s job was to make those two classes separable enough that the arm knows what it is aiming at — and what it must not cut.

The target is not “the leaf.” It is the stem at the base of the leaf, where you actually snip — and it has to sit so the scissors can come down vertically. The live feed from the XIAO Sense puts a mark on that point and a crosshair where the cutter sits in the image. The motors iterate: move a bit, look again, keep the mark on the crosshair. When it holds, the scissors close. That is visual servoing on four axes. Easy to write on a whiteboard. Messy when the camera is cheap, the gears are printed, and the basil moves when you touch it.

The interesting failure modes are the obvious ones: the model confuses stem for leaf, the servo overshoots, the plant is no longer where the last frame said it was. That is the experiment.

In motion

Phone video at VU StartHub, 2025. Same loop from outside: no joystick, printed gears, hobby servos, camera on the wrist. Two takes. The camera’s own view is in Seeing and reaching.

Work with us

Firmware, printed parts, a vision loop that lives with four axes: firmware, CAD, on-device AI, and computer vision. Email services@olivabot.com.

If the question is whether a harvest robot is the right move for a crop, that is consulting — a call and a written report.