Build log
The AI proposes. Tested code decides.
You can connect Lychnos to the AI model you like. That model can ask Alkyon to fly somewhere. It cannot make it fly.
In short. Language models are good at understanding what you want and bad at guarantees. So in Alkyon the model only proposes; a stack of small, tested programs that do not learn decides what the motors do. In simulation, the fastest of them reacts within a fraction of a second, quickly enough to stop the robot at its indoor top speed before an obstacle that appears less than a metre ahead.
Why the line is there
An AI model can misunderstand, be tricked by a message it reads, or simply be wrong. That is acceptable when it drafts an email you will read before sending. It is not acceptable when it moves a robot through a room with people in it. So we split the work: the model turns “go and see if the window is shut” into a task; deterministic code turns the task into motion that keeps every rule, or refuses it.
The layers
| Layer | What it does | How fast |
|---|---|---|
| Thinking | The AI model: understands you, proposes a task | 1 to 10 s |
| Planning | Turns the task into a route on the map of your home | 0.1 to 1 s |
| Reflex | Slows or stops for anything in the path the robot is about to sweep, from live sensor data, with no map | 1 to 10 ms |
| Safety guard | Checks the final command against fixed rules, above everything else | every command |
The rules the guard enforces
- Contact: the motors stop.
- Two presses on the ring, a critical battery, or lost sensors: it lands.
- Speed: at most 1 m/s at home, 0.5 m/s with people in the room.
- People: at least 1 m away, with the margin growing with the robot’s own uncertainty about where it is.
- Places: no-go zones, and a margin from the floor and the ceiling.
The guard takes no advice from the model, and it does not learn. Learned parts may only propose; tested code approves. That rule covers everything we will add later, too.
What the simulation shows
- The reflex decides in a few millionths of a second, and the whole chain, from a new depth frame to the body tilting away, takes a fraction of a second.
- At its indoor top speed, it stops without contact for an obstacle that appears less than a metre ahead.
- Pressing the ring twice stopped the robot within our target time in most of our tests and a little later in the rest, so we are still working on it.
These come from simulation with realistic sensor noise and delays, not from a flying robot. The test stand comes next.
The same rule on your finger
The ring follows it too. Lychnos can draft, suggest and ask; sending, accepting, deleting or buying waits for your press, and paying is not something the AI can do at all. More on that in the AI agent that asks first.