Drone swarm feedback and control
Getting a drone to go where you tell it is harder than it sounds. This is how the swarm does it.
Finn Wilson
Control, machine learning, physics and robotics.
Now
I'm an engineer in Sydney. I trained in mechatronics, space engineering and physics at the University of Sydney.
From October to December 2026 I'm on a research sabbatical, finishing papers on automated drilling, on the cost of keeping a system organised, and on how much a herding drone needs to know about a flock. I'm also working on how to prove tethered drone swarms are safe, whether AI slop is load-bearing, whether coding agents understand code or just recall it, cameras that can spot a drone a kilometre away, how wind feeds ocean waves, and some superconductivity. I'm available from January 2027.
Before this I founded Tacita, where I built AI systems for legal work and long-horizon agent work, and I was a systems engineer at Raytheon Australia.
Selected work
Ferroflock: drones that fly like a flock
A swarm of cheap Tello drones, motion capture and about 30,000 lines of C++. The drones fly as a flock, and you can herd them live in the lab.
When does automated drilling stop learning?
Software that picks drill holes to cut uncertainty spreads them out, and stops collecting the close pairs it needs to learn how fast ore grade changes with distance. The usual scores don't notice. A cheap check after each batch does.
The cost of preserving organisation
Keeping a system organised until a deadline can be free, and keeping it organised afterwards can still cost real performance. We prove when, for a bird flock, a superconductor and a power-grid generator.
Safety guarantees for tethered drone swarms
When several drones are tied to one collar, can you prove their cables won't tangle? I'm building the proof on the braid group, and my own contact solver for the friction values it needs.
Watch
Latest essay
The Electric Motor Problem
Technology pays off only when a system is redesigned around its bottleneck. Firms that bolt AI onto old workflows repeat the electric-motor mistake.