Live view · black box

It teaches itself to land.

Blank network. No reward, no training data. It test-fires, learns what its own vehicle does, and lands. The goal is a question asked at runtime, never written into the weights.

waiting

alt — m · vel —

Flight record

Flights

—

Landings

—

Crashes

—

Validation

—

Plant and learner stay on the host. This tab draws the last altitude, throttle, and imagined trajectory.

Goal: a runtime question

Same weights. Change the target mid-flight; it was never stored in the network.

Pace

Real-time is the default: 10 ticks/s on the vehicle clock, so you can actually read what it is doing. Fast (~300 ticks/s) compresses the learning once you know what you are watching. Guard intercepts: 0

What you’re looking at

The setup

A vehicle with one thruster has to come down and land softly. It begins knowing nothing: not gravity, not how strong its own engine is, not what happens when it fires. There is no reward function, no training set, and no human scoring it.

So it does what an engineer would do with unfamiliar hardware: it test-fires. It tries a throttle, watches what the vehicle actually did, and builds an internal model of its own machine. Then it uses that model to plan.

Reading the stage

The vehicle: altitude and descent rate are live.
The faint trajectory: the future it is imagining right now. Before each command it simulates what would happen, using the model it taught itself.
The envelope: the safe operating band. Any imagined future that breaks it gets refused.

Guard intercepts counts the commands the envelope check threw away before they were ever executed.

What to watch

Test-firing. Early motion looks aimless because it is: it is gathering evidence about its own dynamics. This is the phase that replaces a training pipeline.
Landing. Once the self-taught model is good enough, it plans through it and puts the vehicle down. Nobody ever wrote “how to land” anywhere.
Ask for a hover instead. Type an altitude and send it. The vehicle holds there with the same weights. The goal was never stored in the network; it is a question asked at runtime, so changing your mind costs nothing and needs no retraining.
The guard arms itself only after a validated snapshot of the model passes its own check. An unproven model is not allowed to veto commands. A monitor that hasn’t earned trust is worse than none, because it starves itself of the data it needs.
Real-time vs fast. It opens at the vehicle’s own clock so the motion is readable. Once you know what you are looking at, Fast compresses ~30× and the learning curve becomes obvious in one sitting.

Scope

It crashes while it is learning. That is what test-firing costs, and the counter shows it. The claim is not that it lands perfectly on the first try; it is that competence, the model, and the safety envelope all came from the vehicle’s own experience, with no author, no labels and no cloud, and it could not diverge at any point along the way.

Everything else on this site learns a task someone specified. This one had to work out what its own body does first, then be told what to want at runtime as a question, changeable at any moment without touching the weights.