Floor 2 · 3 exhibits
The Turtle Studio
Type a word, a turtle moves. Say what to draw, and watch the words become actions.
A real, in-browser LOGO: a genuine tokenizer → parser → tree-walking interpreter, plus a transparent natural-language decoder you can read and teach. Then the turtle grows up into a trainable robot pet with real sensors and a 50 Hz sense→decide→act loop. Words → robotics, made literal — the warm on-ramp to coding and AI. And when one robot isn't enough: a deterministic planner that turns a goal into a team of tasks, with a real STOP button.
The Turtle Studio
Say “draw me a square”, watch the words become actions, and a turtle draws it — live.
Type a word, and a turtle obeys. The AI-first, honest answer to Scratch — everything here is genuinely computed, nothing is faked.
How deep would you like to go?
show me · the plain-English storyYour words become actions. The decode panel lights up each word with its job — action, shape, size, colour — then writes the real little program (LOGO) that draws it. Try a challenge, and the turtle checks your actual drawing, not a pretend “well done”.
Wow — Nothing is scripted: your words are decoded into real LOGO by rules you can read, and a genuine tiny interpreter (tokenizer → parser → tree-walking VM) runs it. Change a word and the drawing changes because the parse changed — and any word the turtle doesn't know, it flags rather than fakes.
Try this — Say “draw me a big red star”, watch how each word lights up in the decode, then teach a brand-new word: type “a house is a square with a triangle on top” — now say “draw a house”.
What's really happening · A real, tiny LOGO interpreter built from scratch — a lexer, a recursive-descent parser to an AST, and a tree-walking VM that emits the exact list of line/move operations the canvas draws (and only those). The natural-language decoder is a transparent rule/grammar grounder: it tags every word by role, resolves an inspectable intent, and generates the LOGO you can see — the same word→action mapping a trained language model performs, except here the rules are readable and teachable, and unknown words are flagged, never hallucinated. Deterministic: the same program always draws the same picture. Local-first and free — a child's taught words live only in their own browser, with no account and no tracking.
Read the full lab reportThe Robot Yard
The turtle grew wheels and senses. Now you don't command it — you train a brain, and it decides for itself.
Your turtle grew up into a little pet with motors and eyes. It can't see the whole yard — only feel what's right in front of it. Teach it a brain and watch it decide, all by itself.
How deep would you like to go?
show me · the plain-English storyThe pet can't see the whole yard — it only feels what's right in front of it: how far the wall is, whether there's a line underneath, whether it bumped. Turn on the SENSE→DECIDE→ACT panel and watch the loop: every tick it senses, checks your rules to decide, and moves. When it follows the line, it really is deciding that itself, from what it felt — not replaying a recording.
Wow — Nothing here is scripted or filmed: the path is real differential-drive physics integrated 50 times a second, and every turn is a genuine rulebook read against real sensor readings, tick by tick. Same brain, same yard, same path — every single run. Load “follow the line” and it truly follows it; it decided that itself, from what it sensed.
Try this — Load “Follow the line” and press Run — watch it feel for the line and steer itself to the goal. Then turn the dial to 🧑 and open the SENSE→DECIDE→ACT panel to see exactly why it turned, every tick.
What's really happening · A genuine, from-scratch robot simulator — not a video, not a script. A differential-drive body (two wheel speeds → pose) is integrated with a fixed 20 ms (50 Hz) Euler step, so it is fully deterministic: no randomness, no wall-clock, the same brain + world traces the same path every run. Three real senses are computed from the world each tick: a ray-cast ultrasonic distance to the nearest wall, a line sensor sampling the painted floor under the nose, and a bump flag from real circle-vs-wall collision resolution. Two honest brains drive it: sequential LOGO (the very same tokenizer→parser the Turtle Studio uses, compiled to timed wheel commands) and a reactive EVERY-TICK rulebook (its own real tokenizer→parser→evaluator) — the true sense→decide→act loop. Every challenge is graded on the actual simulated trajectory, never a scripted “well done”. The pet genuinely “decides” by running your readable rules against its senses — it is a real reactive controller, not magic. Local-first and free: a child's brain lives only in their browser, with no account and no tracking.
Read the full lab reportA Director With a Real STOP Button
A deterministic goal→plan-DAG expander that's genuinely interruptible.
Press Run and watch a plan execute step by step — then press STOP, and watch it actually, immediately, stop.
How deep would you like to go?
show me · the plain-English storyA goal expands into a real dependency graph (a plain lookup table, not a language model), then a deterministic state machine walks it tick by tick, running independent branches in parallel. Press STOP at any point — it halts immediately, and the tick counter genuinely stops moving.
Wow — STOP genuinely clears the interval driving execution — the tick counter freezes on the spot, not just visually, and nothing runs again until Run is pressed.
Try this — Pick a goal, press Run, and press STOP halfway through — then press Run again and watch it resume exactly where it left off.
What's really happening · An illustrative, from-scratch reimplementation, not the real production engine (which stays protected): a static task/dependency graph, expanded by a plain table lookup with zero LLM involvement, executed by a pure step() state machine over a pending → running → done graph — the same shape that report names. STOP clears the interval driving execution; nothing runs after that click until Run is pressed again. Ticks are simulated units of work, not measured wall-clock durations.
Read the full lab report