UBRIUniversal Brain-to-Robot Interface

A manifesto on neural oversight of robots

The brain already knows when a robot gets it wrong.

UBRI turns that knowledge into something robots can learn from. We are building a comfortable EEG device, a way to synchronise it with a robot to the millisecond, and the first open dataset of people's brains watching robots work.

Fig. 1  Drift and correction. The oldest image of oversight is the helmsman, kybernḗtēs in Greek, the root of both govern and cybernetics. He does not row. He notices the drift and corrects it. Each red mark is a correction made the moment an error is noticed. UBRI reads that moment from the brain.

§1

Autonomy is outrunning oversight

Robots are leaving the lab. They sort parcels, fold laundry and move through homes, and every one of them still makes mistakes that only a person can spot. Today that person corrects the machine with a joystick, a screen, or a remote operator paid by the hour. Each correction costs attention, so most mistakes go uncorrected and the machine learns nothing from them.

The more capable robots become, the worse this scales. We think the bottleneck of the coming decade is the bandwidth between the people who know what is right and the machines that act.

The autonomy of robots does not guarantee the autonomy of humans.

§2

A signal we already produce

When you watch someone, or something, make a mistake, your brain reacts before you decide to say anything. Within about half a second a characteristic wave appears over the top of the head: the error-related potential, or ErrP. It is involuntary. It needs no words, no movement and no training on your side, and neuroscience has studied it for decades.

Used as feedback, it could tell a robot it got something wrong at the moment it happens, from a person who is simply watching. No button, no annotation session, no teleoperation rig.

§3

The bottleneck is data

Detecting the ErrP in a laboratory works. What does not exist is the data to make it work for anyone, anywhere, with any robot. Raw EEG on its own is close to worthless. It becomes valuable when every millisecond of it is tied to what the person saw and what the robot did: the robot's state, its camera, the task, the instant the error happened.

That anchored data is what we are building. Brain signals synchronised with robot behaviour, at a precision we hold under a millisecond, in a format any robotics lab can open with the tools it already uses.

§4

Hardware built for volume

Medical EEG caps give beautiful signals and take half an hour of gel to fit. Data at scale needs the opposite trade: a device people put on in a minute, wear for an hour and forget about, built from parts that can be made by the thousand. We accept a somewhat noisier signal in exchange for many more hours of it.

That is why we design our own acquisition board. Every sample it sends carries what a dataset needs and most hardware leaves out: its exact timing, a check that it arrived intact, and an honest count of the ones that did not.

§5

Open by default, consent first

A small team cannot collect the world's neural data alone. We publish the data format and the first dataset in the open, in the ecosystem robotics labs already work in, so others can record with their own hardware and add to it.

Brain data is the most personal data there is. Every recording we publish traces back to explicit consent that can be withdrawn, and our export tools refuse any recording without it. A field that handles neural data the way advertising handles clicks would deserve to fail.

§6

What could prove us wrong

We would rather name the weak points ourselves.

  • The ErrP also fires on surprise. A robot doing something unexpected and correct can look like a mistake, and a robot rewarded for never surprising anyone learns the wrong lesson. Our protocols are designed to tell the two apart.
  • Eyes move, people blink. Eye and muscle activity can imitate a brain response. On public data we found that blinks alone could inflate detection scores, so our dataset records what is needed to control for them.
  • Dry electrodes may be too noisy. This is the first thing we will measure on real hardware, against pass and fail thresholds written before the test. We will publish the result either way.
  • Consumer EEG has a graveyard. Many headsets shipped and few stayed in use. We start with robotics labs, who already pay for human feedback, long before any consumer product.

§7

Where we are

Done, November 2026

  • On public EEG data, our detector recognises the ErrP of people it has never seen, without calibrating to them: AUC 0.85 with a research cap, 0.79 with the seven electrodes our device will carry.
  • An open data format for brain signals next to robot data, which accounts exactly for every lost sample.
  • The acquisition chain, from the board's firmware to the dataset, running end to end in simulation.

Next

  • An evaluation bench on real hardware: electrode noise, timing, wireless losses.
  • Our own board, validated against a reference amplifier.
  • The first public recordings of people watching robots make mistakes.

§8

Contact

We are looking for robotics labs that want a human feedback signal, people who know EEG better than we do, and funders who share this thesis.

Enguerand Chrétien, founder. Paris.

hello@getubri.com LinkedIn

Further reading

  1. Chavarriaga R., Millán J. d. R. Learning from EEG error-related potentials in noninvasive brain-computer interfaces. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 2010. The public dataset behind our first results.
  2. Chavarriaga R., Sobolewski A., Millán J. d. R. Errare machinale est: the use of error-related potentials in brain-machine interfaces. Frontiers in Neuroscience, 2014.
  3. Iturrate I. et al. Teaching brain-machine interfaces as an alternative paradigm to neuroprosthetics control. Scientific Reports, 2015.
  4. Salazar-Gomez A. F. et al. Correcting robot mistakes in real time using EEG signals. ICRA, 2017.