// Work
What we build.
Three lines of work. Coordination, vision, simulation.
// Coordination
Decentralized multi-agent coordination
Swarm architectures with no command node. Each agent follows local rules and reacts to
what it can sense nearby; the useful behavior — coverage, search, regrouping after
losses — is not written into any individual agent, but comes out of the interaction
between them. It runs in simulation today; the work now is tuning it for specific
missions.
The swarm gets bigger; the problem does not. Nothing is keeping track of the
whole, so there is nothing to overload.
There is no plan to capture and no controller to destroy. Jamming, attrition or a
severed link changes what the group can do instead of stopping it. That robustness is also what
makes the behavior hard to verify.
// Perception
Argus — vision systems
Argus is our line of vision systems: inspection and identification — knowing whether a
thing is what it should be. Defects, parts, equipment. It is built to be trained by the
people who use it, on whatever they need it to recognize; there is no fixed class list.
Training takes images of the item you are checking, in each state that matters.
Argus Station, the fixed installation, is deployed and in live beta testing. It
runs on Android today; ONNX underneath means other hardware is a port, not a rewrite.
Vision is not exclusive to autonomy. The same instrument applies anywhere something
has to see reliably and act on what it sees.
Our first vision work was a volunteer open-source effort identifying military vehicles.
// Simulation
Deterministic simulation and training
A physics and world model that returns the same result from the same seed on every run —
for a behavior, and for the run that trained it. Our own coordination work runs on it. A
behavior can be replayed and inspected by someone who did not write it, which is a
different standard from a recording that happened to go well. The same standard applies
to autonomy claims that are not ours.
Emergent behavior cannot be checked by reading the source, because it is not
written in any one agent. It has to be run.
Every simulated run is recorded in full and pinned by a SHA-256 committed with the code. A
change that moves any number in a trajectory breaks that check, so a physics change has to
be accepted deliberately rather than discovered later.