Simulation
Watching for hornets
Total Buzzkill.
A turret that learns to defend Malta’s beehives from the Oriental hornet. Designed with AI, trained on a phone, tested in simulation.
01 · The threat
Malta’s bees are being hunted.
The Oriental hornet (Vespa orientalis) hovers at the hive entrance and snatches bees as they fly home.
59.5%
of surveyed Maltese beekeepers lost colonies to hornets in 2024. Double the year before.
Survey of 37 beekeepers, The Maltese Herald, 29 Sep 2026
- ~70% of local hives lost in 2022 Times of Malta, 18 Oct 2022
- 30 hives lost at one apiary, about €30,000 Birżebbuġa, Malta News Agency via Italpress, 12 Oct 2026
- Maltese
- il-bagħal taż-żunżan
- Size
- Workers up to 23 mm, queens up to 35 mm
- Season
- Late April to mid-November
Sources: The Maltese Herald (name); Mississippi State University Extension (size); University of Malta, 2021 (season).
02 · The turret
Designed by talking to an AI.
We described it. Claude drove FreeCAD and modelled it: 23 printed parts merged into 14, the main arm shaped by topology optimisation and stress-checked.
Simulation
- Aim
- 90° side to side, 90° up and down
- Eyes
- Three cameras
- Reach
- About 5 m
- Fires
- Biodegradable gel balls, over 98% water
Safety lock on
Drawn to the millimetre.
Assembly sheets generated straight from the CAD model: every part numbered, the pan bearings cut open.
03 · The simulator
A whole apiary in your pocket.
Hives, bees, hornets and passers-by in 3D, with the real turret on guard. Fully offline.
Watch it learn.
Same hornets, two brains: untrained, and after 23,800 practice rounds. One run, not an average.
Scroll to play
Frozen in mid-air.
Slowed 100×: the gel ball leaves the barrel and stops a breath from the hornet.
Scroll to play
Simulation
7.3 mm ball · 52 m/s · 45 ms to the hornet
The app · Android
Get it by QR code at a talk.
04 · Training
It teaches itself to aim.
Nobody codes the aim. A small neural network tries, misses and improves. Train one on your own phone.
23,800
practice rounds behind the demo model
- Sees
- 444 numbers per step
- Thinks
- 256 neurons
- Decides
- Pan, tilt and fire: 16 outputs
- Learns by
- PPO: trial and error
Share a brain by QR code.
- ~11 s screen to camera, no internet Phone screen to laptop webcam
- ~41 KB the whole model
It thought bees were birds.
Off the shelf, a detector called bees “bird”. After 33 minutes of training on mostly Maltese footage, it never confidently mistook a bee for a hornet.
- 0.83 hornet detection score
- 0.90 bee detection score
- 54 ms per frame, laptop CPU
AP@0.5 on footage it never trained on
05 · Safety
People in view? It won’t fire.
A hard lock blocks firing whenever a person is in view. It’s a rule, not something the AI learns.
- PeopleFiring locked.Locked
- LasersReflections could hurt the bees.Rejected
- HardwareSimulation only, for now.Simulation
What it is
Every tool behind it, from the AI that designed the turret to the phone that trains it, is one you can use today.
What it isn’t
A product. Beekeepers are losing hives, and we won’t overpromise.
06 · Learn more
See it live.
More dates coming.
- Talk
- Workshop




