Three drones lifted off a courtyard in Manchester in late September. The pilot stood with his hands in his pockets. No screen, no joystick, no voice link — just a white headband with small pins pressed against the back of his skull, reading electrical activity over the visual cortex and pushing commands to the aircraft over Bluetooth.
The flight worked. NeuroDoctrine, a defence and space startup founded this year, used its own two-day summit to show an operator steering three small commercial drones with neural signals alone. Founder and chief executive Zied Tayeb says calibrating the headband to a new person takes about six hours, because every brain's signal patterns differ.
The more interesting number is the one that did not appear. NeuroDoctrine's website claims 98% accuracy, latency under 300 milliseconds and a swarm ratio of more than 500 systems per operator. The company has published no test data behind any of those figures.
The case for a hand on the loop
NeuroDoctrine's argument starts with a constraint the autonomy industry rarely says out loud: a machine that selects targets still needs a person who can be held responsible for the selection. Tayeb puts it plainly. Fully autonomous systems cannot be trusted with weapons. His second reason is operational. Robots do not learn in real time on a battlefield. People do.
That is a defensible product thesis, and it lines up with policy. The UK Ministry of Defence's 2022 doctrine requires context-appropriate human involvement in weapons that identify, select and attack targets. A neural command layer that keeps a human in the decision path is easier to argue for than a fully autonomous one, at least on paper.
The company has recruited people who know how that procurement conversation works. Lisa Sanders, formerly the US Special Operations Command's director of science and technology, and retired Lieutenant General Roger Cloutier Jr., who commanded NATO Allied Land Command until 2022, are listed as advisers and investors. Steve Furber, an Arm founder and neuromorphic-computing researcher, appears on the leadership page.
✔ Pole A — what the demonstration actually proves
+ Non-invasive hardware avoids implant regulation
+ A software layer can sit on existing drones and munitions
Confirmation criteria: a named defence buyer runs an independent trial and publishes the results.
The file that is not there
The case against is not that the demo was fake. The drones flew. The problem is the distance between a courtyard and a procurement file. On that distance, NeuroDoctrine has published almost nothing.
The site's performance claims have no supporting release, no third-party test and no stated sample size. Nobody outside the company has said how many operators have used the system. The company has not disclosed how much it has raised, from whom beyond its two named backers, or at what valuation. It has not named a customer, a trial partner or a government programme.
The corporate footprint is equally thin. No UK company named NeuroDoctrine appears on the Companies House register. MyelinZ Ltd, incorporated in Manchester in December 2024, is the related entity named in coverage, with co-founder and chief operating officer Samaher Garbaya listed as company secretary. The website advertises offices in Los Angeles, Manchester, Leuven and Luxembourg.
Then there is the part that matters most for a fielded system. Six hours of calibration per operator is a laboratory number. A dry-electrode headband also picks up more movement interference than a gel-based cap, and visual fatigue degrades the signal over a long session. NeuroDoctrine has not said whether its decoding relies on flickering visual targets, a method that is accurate but tiring.
No published accuracy or latency test. No named customer or trial partner. No disclosed round or valuation. No UK legal entity under the NeuroDoctrine name. Four gaps a defence buyer would close before a pilot, not after.
We can literally record, process and interpret complex, noisy brain signals in real time.— Zied Tayeb, founder and CEO, NeuroDoctrine
A decade of demos, a thin product shelf
Brain-signal control of robots is not new. It has been demonstrated in research for at least a decade, and the pattern is consistent: the decoding works in a lab, the calibration is fragile, and the product rarely leaves the building.
| Demonstration | Interface | Scale shown | Calibration | Published result |
|---|---|---|---|---|
| Arizona State University, 2016 | 128-electrode cap | Up to four small robots | Daily | ✔ Research paper |
| University of Technology Sydney, 2023 | Dry graphene headset | One quadruped | Per session | ✔ Up to 94% accuracy |
| NeuroDoctrine, 2026 | Dry-pin headband, visual cortex | Three drones; ten flown together | About six hours | ✗ None |
Demonstration records from published accounts, 2016–2026
Read the table down the right-hand column. The research demos shipped a number. The commercial demo shipped a claim. That is the gap an investor is being asked to underwrite.
Setup time, not accuracy, is the field constraint
Tayeb's own estimate for adapting the headband to a new user. National Defense, 2026
The pattern has a precedent in the lab next door. As we wrote in October, a fruit-fly connectome began driving a simulated body — the wiring diagram was the breakthrough, and the behaviour it produced was the proof. Neural decoding is a research result. A fielded control system is a calibration and trust problem wearing the same hardware.
The trust budget
Assistive neurotech tolerates errors. A prosthetic hand that misreads intent is annoying; the user tries again. A weapon that misreads intent is a different category of failure, and the tolerance for a false positive is close to zero.
That asymmetry explains why the latency claim matters less than it first appears. Under 300 milliseconds is roughly the pace of a human reflex, and it is impressive for an EEG pipeline. The binding constraint is not how fast a command arrives. It is whether the operator can predict when the system will misread one.
NeuroDoctrine's architecture tries to manage that risk by narrowing the command surface. The operator does not fly the drone stick-and-rudder. He triggers a pre-programmed task and lets the aircraft's own autonomy execute it. That keeps a human in the decision path without asking neuroscience to replace piloting skill.
The trade-off is that the product becomes a command selector, not a control channel. Its value scales with the number of pre-programmed behaviours a defence customer is willing to trust. That number is set by doctrine and liability, not by decoder accuracy.
For an investor, that reframes the diligence question. The headline is a swarm ratio of 1:500. The real question is how many independent decisions one operator can be accountable for in a minute, and what happens to the chain of responsibility when the headband fires the wrong one.
What has to be true by 2028
Four things would move NeuroDoctrine from a courtyard to a contract. Each is testable, and none requires a new scientific breakthrough.
First, an independent test with a stated sample and a stated latency distribution, not a marketing percentage. Second, calibration that a soldier can complete in under thirty minutes, ideally from a cold start in the field. Third, a named defence ministry or prime contractor running a paid pilot with published milestones. Fourth, a disclosed cap table and a legal entity that a procurement officer can contract with.
Until then, the honest reading is that NeuroDoctrine has built a credible demo and a plausible thesis on top of a decade-old research base. The human-in-the-loop argument is stronger than the autonomy argument for weapons. The evidence for the product is weaker than the evidence for the principle.
The number to watch is six hours of calibration, not 98% accuracy.
A named trial partner would be worth more than another summit.
The headband worked in a courtyard. The operating system is still a slide.