What our ring and robot learn, and what they may not.

The rule first

In our design, a learned part only proposes. Plain, tested code checks every proposal before anything moves, and the safety guard at the top, the code that keeps the distances and stops the motors on contact, does not learn at all. Every model below lives under that rule: the AI proposes, the code decides.

1. A camera that knows four things

A small model, trained from scratch on openly licensed photos, spots people, curtains, cats and dogs. Its second version saw people better and lost some of the tulle, so we did not use it; a third, trained more evenly, matched the first. It is not in the product: it has not yet looked through the robot’s own camera. The whole story: teaching a robot to see cats, dogs and tulle.

2. A flight controller tuned on a million simulated robots

A flight controller has a handful of numbers that decide how hard it corrects. Instead of tuning them by hand, we let a search try them on a graphics card, tens of thousands of simulated robots at a time, each slightly different: motors a little slower or faster, the body a little heavier or lighter, its weight a few millimetres off centre, the arms in any of their shapes. Then we tested the winner on more than a million conditions the search had not seen, with sensor noise. With the old numbers, the robot would often have needed more room to stop than its reflex allows for; with the new ones, in none of them. Of everything here, it is the only part our flight code uses today, and what it learned is a set of numbers that plain code applies.

3. A guess at where to land

Landing is strict work. The robot measures the surface below it with a downward distance sensor and lands only where plain rules say the pad fits, the surface is level enough and the guards have room. Often the rules find nothing in view, even when a good spot is close: how our robot chooses where to land. So we trained a small network on thousands of looks from simulated landings to guess where a landable spot is likely. In tests, its best guess was usually right, including in many views where the stricter rules had found nothing. It does not land the robot. It only says where to fly and look next, and the rules check again. It stays a candidate until a campaign shows it lands more often with no unsafe outcome.

4. A speech model that learned to invent words

The ring hears you from your hand, through whatever noise is around. We have made a series of attempts to adapt open speech-recognition models to that sound, in simulated cafés, streets and trains. Each attempt is now judged by rules written down before it runs: no scene worse, no more invented words, better on what passes the ring’s own checks, and no worse on clean speech.

One version understood our practice scenes better, and then, where the voice was buried under noise, began to write words nobody had said, far more often than the model as released. The reason was almost funny. Every clip it trained on had words to write, so it learned to always write some. None of the attempts has been good enough to use yet, so the open model as released is still our candidate. Earlier results: can a ring understand a mumble?

What learns from what

Every simulated trial is recorded: what the robot saw, what it did and what happened. Those records taught the landing network, and they are what the next models will learn from. Learning from your own home is a different matter: we plan to do it only on the dock, in your home, and only with your consent.

What this does not prove

All of it comes from simulation and public data. Nothing here has been measured on a real ring or a real robot, and each model will be tested again with the hardware’s own sensors before it is used. Lychnos Ring is in development; Alkyon is a concept that flies only in simulation.

Source for the search method, read 4 November 2026: Rubinstein and Kroese, a tutorial introduction to the cross-entropy method, in The Cross-Entropy Method, Springer, 2004. The results are our own simulation tests.