We bring the brain into your system.
The plan is to take the brain into partners’ own host environments — their drones, ground vehicles, simulators and test ranges — and run it there, against their conditions instead of ours. We write the adapter; your system stays yours.
Your failures become its rules. Failed-run logs and incident telemetry are what the brain learns from. It turns what went wrong on the last system into rules for the next one, including weapons-capable platforms under human command. We have done it once from a public record: 8,219 US federal crash and incident reports on automated vehicles became 15 driving rules, 6 of them enforced so far.
No partner integration exists yet. This is the next step, not a result.
Milestone 05 · The ring: why a robot boxing ring before a battlefield?
Because the ring is the cheapest place to find out what breaks when the other side hits back. Perslis plans to buy a commercially available humanoid fighting robot and enter sanctioned robot-vs-robot bouts, with the remote pilot replaced by the brain. We do not build the robot hardware. It is the proving ground on the road to combat-capable humanoids under human command — and where we raise the money to get there.
- Robot sensors
- Eye
- Peel
- Floor
- Robot actuators
A human keeps the stop button and safety authority the whole time, but does not steer the robot during an autonomous run. The vendor’s own safety systems stay on.
Why not just test it in the lab?
In the lab, we control the fight. A competition brings independently developed opponents, real contact physics, balance disruption, occlusion, sensor degradation, hardware wear, unpredictable behaviour, and situations the Perslis team did not script.
The score isn’t just “did it win?”
- Human interventions per autonomous hour
- Recovery rate
- Novel failure classes
- Floor vetoes
- Hypotheses the system generated itself
- Rules promoted and rejected
- Reduction in repeated failures
- Transfer of learned symbolic knowledge to opponents it has never seen
Losses stay on the public record. A lost match is evidence, not automatically a failed benchmark: every impact goes down the same chain — state trace, hypothesis, controlled change, retest, reject / not proven / promote.
Earned, not bought first
The expensive humanoid is bought only after the earlier milestones pass their gates: hardware-in-the-loop, low-cost physical platforms, then the humanoid testbed in our own lab. The gates are written before the run, like every Perslis experiment card.
Where the field is today
Humanoid bouts today are mostly piloted by people — VR rigs and gamepads — with short autonomous sequences; a robot in the Unitree G1 class starts at about US$13,500, research configurations cost more. The vendor has also shown its own learned model sparring without a pilot. Our question is different: not whether a robot can fight on its own, but whether a brain that learns without a neural network, under a floor, with a trace for every decision, survives contact.
- Interesting Engineering, humanoid bouts at CES 2026 ↗
- eWeek, CMG 2026 humanoid combat competition teaser ↗
- TechEBlog, G1 sparring with the vendor’s own learned model ↗
- RoboZaps, Unitree G1 price and specifications ↗
Sanctioned robot-vs-robot bouts only. No bouts against people, and no hardware development: we put our brain in a robot someone else built.
The robot is not the product. It is the instrument. The brain is what is being tested.
We know what this architecture does when we control the fight. Seed funding lets us stop controlling it, one step at a time.