Questions, answered.

Not yet in this preview. The cultures you can train on, the sandbox culture and the CL1 device twin, are simulated: a model of a dissociated cortical culture built from published values, including a simulated response to training. The remote CL1 shows how a rented device is reached, but no hardware is connected.

The one real source is a recording of rat cortical neurons from Wagenaar, Pine and Potter (2006, BMC Neuroscience). You can watch it drive a robot, but it cannot be stimulated, so it cannot learn. Compare the biocomputers.

Watch-only behaviours, which read activity without stimulating. Look toward activity turns the Unitree G1 toward whichever half of the electrode array is busier. Training needs stimulation, so trainable skills run on the simulated cultures.

Every validation runs the skill’s test three times on the same spontaneous activity: with stimulation, with no stimulation, and with the neurons’ spikes blocked from the decoder. A skill counts only if it passes with stimulation and fails under both controls.

If it also passes without the neurons, it is marked invalid: the motion does not depend on them. How validation works.

One JSON file in the robooi.policy/v0 format: the skill file itself, the decoder wiring found in calibration (for each pathway, the electrode pulsed, the electrodes read and the ON threshold), the learning record (episodes, success over the last five, the feedback protocol) and the validation result for each condition.

Its provenance says which neurons trained it: in this preview, a simulated culture. You can download it once a run is validated.

Three phases in one closed loop. Calibrate: every electrode is probed with single pulses to find the sites the culture answers on. Train: episodes of the skill’s test, each followed by feedback. Validate: the test against two negative controls. In the sandbox a run takes about 1–2 min. The full walkthrough.

The feedback rule from Kagan et al., “In vitro neurons learn and exhibit sentience when embodied in a simulated game-world” (Neuron, 2022). After a successful episode 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.

Five, all in real MuJoCo physics: the Unitree G1 humanoid, the Franka Panda robot arm, the Unitree Go2 quadruped, the Shadow Hand and the Skydio X2 drone. Each comes with ready skills, and you can add your own from demonstrations.

Yes. Upload a CSV or JSON of joint positions recorded on your robot (teleoperation, motion capture or a scripted controller): a time column plus one column per joint, named like the robot’s controls, with at least 20 frames. Every robot has a sample file to start from.

Poses the robot holds still for at least 0.25 s become targets; the first one is rest. Each target gets its own pathway and a test that cues it.

Only you. Every account gets its own lab: its own sandbox culture, runs, trials, recordings and skills. The server returns only your own records, the live stream carries only your workspace, and skills you create from demonstrations are visible only to you.

On the shared devices the culture itself is shared, which is why each takes one customer at a time. Passwords are stored as salted scrypt hashes, and sessions use an HTTP-only cookie.

The CL1 device twin and the remote CL1 are shared cultures, used by one customer at a time. If one is busy, your run joins a queue and starts automatically when it frees up, so you can close the tab. Queued runs survive a server restart, in launch order. Your sandbox culture is never shared.

No. The robot physics and the cultures run on our servers; your browser draws the robot and the live electrode traces.