The decision layer
At every step, Automata Room describes only the part of the room the mission needs, lists what the robot could legally do next, and lets a decider choose. The simulator carries it out, or refuses. Nothing here is a claim you have to take on trust — the loop is below, and it runs in your browser.
Running here, now
Pick a room, a robot and a mission. The rule-based decider takes it from there: it needs no key, no model and no network call, and every decision it makes is kept for you to open.
Press Run. Each decision below is one pass of the loop: the state it was given, everything it could legally do, and what it chose.
Four working parts
Not cleverness. A state small enough to reason about, a menu that contains nothing impossible, a way to say “not sure”, and a room that answers back.
A delivery is described as its cargo, its destination and what is already done. A patrol is described as its stops and which are still to make. The rest of the room is not in the state at all: it would be noise, and noise is what a decider chokes on.
The options at each step are enumerated by the simulator from the robot's body and the room as it stands. A quadruped is never offered a lift. A few legal but useless moves are always in the list, because a menu with one right answer is not a decision.
A model answers with a choice and how sure it was. Below 50% the step is taken by the rule-based layer instead, and the run records that it was handed back. A confident-sounding guess is the one thing a robot must never act on.
When a route is blocked or there is no floor left to set something down, the refusal goes into the next state and that target is set aside. A delivery that rearranges the room gives it another go. The plan follows the room, not the other way round.
Two layers, one loop
The deterministic layer is not a fallback that sits in a drawer. It runs the loop on its own, it is the baseline every model is scored against, and it takes over the moment a model is unsure or unreachable.
Nearest thing that still needs doing, every time. Instant, free, no key and no network call — the same engine that plans routes and refuses the impossible. Up to 30 decisions a run.
The same state, the same menu, answered by a model, on Automata Room's key or your own: a choice, a probability for every option, and a confidence. Everything it decides is kept — what was offered, what it took, how sure it was, and whether it actually helped.
Jev decides from a free daily allowance, or on a TypeSafe key of your own; the rule-based side runs for nothing, for ever.
Where the edges are
This is the part every page like this leaves out, so it goes here, in the same size type as the rest.
No cameras, no depth, no SLAM, no detector. The scene graph is given by the studio: objects are where the layout engine put them, named by the catalog. Nothing here is seen.
Routes are planned on a 0.25 m grid around object footprints; reach and carrying are rule-based. Nothing falls, slips, or collides with force. It is a kinematic preview.
The models are stylised, built to each robot's real proportions so clearances read correctly. Good for task planning, and for watching a decider work. Not a substitute for a simulator that models forces, or for hardware.
In the studio the same loop runs with your own model, side by side with the rule-based baseline. The whole run downloads twice over: as one documented JSON file, and as labelled rows a notebook can score a model on or train one with — the state at every step, everything that was on offer, and which of it actually helped.