SWUM · SWARM COORDINATION FOR PHYSICAL AI
Many robots.
One goal.
A fleet of robots splits a goal, agrees who does what, does each job once, proves it and learns from it. No controller, and no reliable network required.
HOW IT WORKS ↓33/33
NUMERICAL CHECKS PASSING
13
PROVED RESULTS
6 GB
ALL IT NEEDS ON AN 8 GB MODULE
1
RADIO FRAME PER BEACON
THE PROBLEM
Lose the controller or the link, and the whole fleet waits.
Every fleet today hangs on one controller and one radio link. When either fails, every robot stops. Nothing degrades gracefully.
And when the work is done, the only proof is the vendor's own log. No one else can check it.
HOW IT WORKS
Five steps, each one checkable.
Every robot broadcasts a small authenticated frame that carries what it knows: its neighbors, which jobs are done, and who claims which job. Robots merge what they hear, so the swarm converges to one view no matter how frames are delayed, repeated or reordered.
WHY IT HOLDS
The guarantees are theorems, not promises.
Exactly one robot starts each job (Theorem 4). The swarm converges to one shared view on any lossy radio (Theorem 1). Integer inference gives the same answer on every chip, bit for bit (Proposition 7).
Thirty-three numerical checks cover every theorem, bound and wire constant, and all of them pass. The first draft of the start rule failed those checks, and the rule was fixed before any code was written.
PROOFS & SPEC →ALREADY RUNNING
Built on rails that are already live.
FOCUS
One product, three markets, in this order.
MILESTONES · MONTHS AFTER WE RAISE
From a Gazebo prototype to a field pilot.
Now
EARLY PROTOTYPE IN GAZEBO
~6 mo
THREE REAL ROBOTS
~7 mo
FIELD PILOT, 5 TO 10 ROBOTS
WHO BUILT IT
Two founders. No outside capital.
One of us wrote SUM Chain and built RUMUS, our Rust training framework. The other has shipped more than thirty data products to over a thousand users.
One hundred percent founder owned.
