From a robot to a validated policy.

Pick a robot.

Five robots run in real MuJoCo physics on the server: the Unitree G1 humanoid, the Franka Panda robot arm, the Unitree Go2 quadruped, the Shadow Hand and the Skydio X2 drone. You see exactly what the neurons make them do.

Choose a skill, or upload demonstrations.

Start from one of 14 ready skills, or drop a CSV or JSON of joint positions: a time column plus one column per joint, named like the robot’s controls, with at least 20 frames.

Recorded by teleoperation, motion capture or a scripted controller, the demonstration is resampled to 50 Hz. Every pose the robot holds still for at least 0.25 s is found; the first becomes rest and up to five others become targets. Each target gets its own pathway, and the test cues them in turn.

Rent living neurons.

A biocomputer is a culture of neurons on a multi-electrode array: 59 electrodes, 200 µm apart. The loop runs on it continuously: stimulate, record, feedback every 10 ms. In this preview the cultures you can train on are simulated; the real neurons are a recording you can watch.

Compare biocomputers →

Calibrate.

Before training, each of the 59 electrodes gets 6 single pulses, 60 ms apart, and the platform listens for 30 ms after each, ignoring network bursts. Each pathway is placed on the stimulation site whose nearby electrodes answer most reliably, at least 400 µm from the other pathways.

Its ON threshold sits 3.5 Poisson standard deviations above the read electrodes’ resting rate and never below 0.85 spikes per electrode per 100 ms, so the response itself has to grow during training.

Train with feedback.

Each episode replays the skill’s test: its pathways are cued on schedule and the robot moves only if the neurons answer. The checks then score the episode.

After a success the culture gets predictable stimulation: 100 ms at 100 Hz on the pathways’ own electrodes. After a miss it gets 4 s of unpredictable pulses at 5 Hz on random electrodes, then a 4 s pause. The protocol follows DishBrain (Kagan et al., Neuron 2022).

After a success

Predictable: 100 ms at 100 Hz, same electrodes

After a miss

Unpredictable: 4 s at 5 Hz on random electrodes, then a 4 s pause

Ready to validate

At least 8 episodes, 4 of the last 5 successful and the last 3 all successful

Budget

40 episodes per session

How the culture responds to feedback is simulated in this preview.

Validate against two negative controls.

Validation runs the test three times on the same spontaneous activity: with stimulation, with no stimulation, and with the neurons’ spikes blocked from the decoder. It passes only if the checks pass with stimulation and fail under both controls.

Passed normally and failed under both controls: the neurons caused the motion.Verdict: pass

Also passed without the neurons: the motion does not depend on them.Verdict: invalid

Download the policy.

One JSON file: skill, decoder wiring, learning record, validation.

Format

robooi.policy/v0

Skill

The skill file that was trained

Decoder

The calibrated wiring of each pathway: electrode pulsed, electrodes read, ON threshold

Learning record

Episodes, success over the last five, the feedback protocol

Validation

The verdict and the result of each condition

Provenance

Which neurons trained it: in this preview, a simulated culture

The method behind the feedback.

DishBrain embedded cortical neurons on a multi-electrode array in a simulated game of Pong. A hit was followed by predictable stimulation and a miss by unpredictable stimulation, and the cultures improved their play within minutes. RoboOI applies the same feedback rule to robot skills.