From a toy’s brain to a combat humanoid under human command.
Learn in simulation. Transfer lessons forward. Validate in the physical world. Only the first step is done.
Nothing here has been in the field. It has been tested in games and simulation. Here is what that proves and what it does not.
Where it came from
It started as a small brain for a toy that had to work offline. That is still the rule: it thinks on the machine, not in a data centre. Every step since has asked the same question at a larger scale: how do you build a brain you actually own — one you can open, read and fix?
- Day 1The brain factory
- Feb–Mar 2026TinkyBrains
- Mar–May 2026SML — the smallest language model
- Mar 2026 →Structured knowledge
- Mar 2026 →Trace — Peel
- Sep 2026Fail-First
- Sep 2026Peel — a self-evolving symbolic brain
- Sep 2026One brain, many worlds
- NextThe embodied brain
The road ahead
- 01running
Games and simulation
Doom, Wolfenstein, Fallout, GoldenEye, tanks, a drone flight stack in software, photoreal driving. This is where every number on this page comes from.
crutch it takes awayNothing yet: perfect resets, the game telling it the truth, opponents it already knows.
- 02next
Real controller, simulated world
The same brain on a real flight controller and motor driver; the world is still simulated.
crutch it takes awayPerfect timing: real latency, a real bus, a real clock.
- 03gated
Cheap robots in the dirt
Small ground and air robots, where a crash costs a part, not a program.
crutch it takes awayThe game telling it the truth: the eye has only its own camera.
- 04gated
Humanoid in our own lab
A commercially available humanoid, on our bench. It never spars against people.
crutch it takes awayCheap consequences: balance, falls and wear are real.
- 05gated
The ring: robot vs robot
Sanctioned robot-vs-robot bouts, the remote pilot replaced by the brain. This is where we raise the money and find out what breaks when the other side hits back.
crutch it takes awayKnown opponents: now the other side controls half the experiment.
- 06gated
Someone else tests it
Someone else runs the protocol, on their robot, against their opponents.
crutch it takes awayUs: the people who wrote the experiment.
- 07gated
Combat-capable humanoid, under human command
The goal. A combat-capable humanoid operating under human command: people hold authority over the use of force, the floor holds the machine to the constraints it was given, and every decision is on the record. Not built. Gated on every step above.
crutch it takes awayThe ring: rules, a referee, and a stop button someone else controls.
The goal: a combat-capable humanoid that answers to you
Long-term validation includes combat-capable humanoid platforms under human command. Perslis is being developed so that learning may change strategy and operational knowledge without silently expanding the system’s authority.
US policy is clear about who decides. DoD Directive 3000.09, Autonomy in Weapon Systems, requires autonomous and semi-autonomous weapon systems to be designed so commanders and operators can exercise “appropriate levels of human judgment over the use of force”. It also calls for realistic testing, system status an operator can understand, clear activation and deactivation procedures, safety mechanisms, and use consistent with applicable law and ROE. Read the directive ↗
That split is the design. Learning changes tactics and what the brain knows about the ground. It never changes what the machine is authorised to do, and every decision leaves a record a commander can review.
That it is built, that it has been in combat, that it complies with 3000.09 today, or that anything learned in a game carries to a battlefield until each step above proves it. Perslis is not a certified weapons-policy implementation.