▶ Try it live: https://punnerud.github.io/MPErubiks/ (camera scanning works right in the browser — phones included).
A visual-first Rubik's cube app in Rust, running in the browser via WebAssembly (GPU-rendered with wgpu): scan your real cube with the camera, watch it in 3D, and follow a step-by-step guide — then train the standard algorithm sets with a timer and per-algorithm scores. Once you know an algorithm, the solver prefers solutions that use it and can even stop there so you execute it from memory.
Designed so children who can't read yet can use it: big colored buttons, painter-drawn icons, animations and color cues carry the meaning; text is always secondary. English/Norwegian with a flag switch, dark/light theme.
Screenshots
All images are the app's live screenshot tests
(crates/cube-app/tests/snapshots/) — they regenerate with the code, so
this is what the app looks like right now.
| Menu (dark / light) | Solve guide | Solved! |
|---|---|---|
  |
 |
 |
| Practice-stop | Settings | Training cards |
|---|---|---|
 |
 |
 |
| Lesson player | Algorithm demo | Practice unlocked |
|---|---|---|
 |
 |
 |
Features
- 📷 Camera scanning — auto-aligning grid (position and scale), per-pixel Oklab hue voting, evidence averaged over a hold-steady window, and a backtracking constraint resolver that fills uncertain cells using the cube's hard rules (9 per color, every piece once, solvability). The result is rotated + relabeled so every color sits on its standard face.
- 🧭 Solve guide — near-instant solutions (≤23 moves; a shallow pre-pass returns truly short ones for nearly-solved cubes), stepped one move at a time with karaoke notation, two-stage double turns, and a confetti cannon.
- ⚙️ Practice-stop mode — the guide halts where a chosen algorithm applies, shows only its name ("★ T-Perm — you know this one!"), and you execute it from memory; "I did it" fast-forwards, "Show me" reveals.
- 🎓 Lessons — beginner steps (daisy → cross → corners → second layer → top) with a steppable demo player and Practice buttons into the drill loop.
- 💪 Training — PLL · OLL · F2L · beginner cards with demo player, drill timer, success/fail stats and a persistent trained-set (MPEdb).
- 🎲 Practice scrambles — pick the algorithms you are drilling (gear in Play): the shuffle then produces a scramble whose solution actually uses them, and a card reports how many moves it is and what share of it is your algorithm.
- 🌍 31 languages — Nordic, most of Europe, and Chinese (Simplified and Traditional), Japanese, Korean, Vietnamese. Tap the flag for a scrollable picker. CJK scripts arrive as font subsets fetched on demand (38-94 KB, containing only the characters this UI uses), so the download stays small no matter how many scripts are supported.
See INNOVATIONS.md for what this repo invented, borrowed, and moved somewhere unusual — including the measured negative results.
Layout
| crate | what |
|---|---|
cube-core |
dependency-free cube model: 54-facelet state, moves derived from integer sticker geometry, notation parser, case recognizer (patterns derived from algorithm inverses) |
cube-solver |
kewb (Kociemba two-phase) integration, scan constraint resolver, hint/guided-solution engine, packed-table codec |
cube-vision |
dependency-free Oklab color voting, grid auto-fit, field-data eval harness |
cube-render |
wgpu: instanced cubies, procedural sticker shader, move animation, offscreen depth + blit into egui, tap picking |
cube-store |
MPEdb storage (file+WAL native, in-memory + localStorage dump on wasm) |
cube-app |
eframe/egui app: menu, play, scan, solve guide, trainer, lessons, settings |
xtask |
gen-table, pack-table, benchmarks |
assets/algorithms.json holds 127 validated cases — every algorithm is
checked mechanically (its inverse must produce a state of the claimed kind,
and no two cases may be AUF-indistinguishable), so data errors fail the
build. assets/table.pack is the solver table the app ships: nibble-packed
pruning depths + route-compressed move tables (per-column deltas along the
coordinate order — the enumeration is already the optimal "route") +
per-section deflate — 0.9 MB instead of 6.8 MB (13 %), unpacking in ~15 ms.
Targets & build
- Web (primary):
trunk serve→ WebGPU with automatic WebGL2 fallback. Camera scanning needs HTTPS or localhost (trunk serve --tlsfor phones). - Native desktop:
cargo run -p cube-app(same UI, manual cube entry instead of camera).
cargo run -p xtask -- gen-table # once: generates assets/table.bin cargo run -p xtask -- pack-table # packs it into assets/table.pack trunk serve # browser dev build
Tests
cargo test --workspace # core properties, recognizer, vision, solver, store cargo test -p cube-solver -- --ignored # 1000 random solves within the bound UPDATE_SNAPSHOTS=true cargo test -p cube-app --test screenshots # headless UI PNGs RUBIKS_SCANS_DIR=... cargo test -p cube-vision --test eval_scans # field eval (dataset outside the repo)
The screenshot suite renders real app screens through wgpu (lavapipe works)
into crates/cube-app/tests/snapshots/. Scan telemetry (cell crops) uploads
to a private dev server for offline tuning; personal photos are never
committed to this repo.
License
PolyForm Noncommercial 1.0.0: free for everyone to use, modify and share for any noncommercial purpose. All commercial rights remain with the copyright owner (Morten Punnerud-Engelstad) — only the owner may commercialize this software or license it onward for commercial use. Third-party dependencies keep their own licenses.