MOCHI Kids — Shape Detective

A kid-friendly (4–6 year olds) adaptation of the MOCHI 3D-shape oddity benchmark (Bonnen et al., NeurIPS 2024 D&B). Three-image trials: two views of the same object, one different — tap the odd one out.

Live demo: https://vislearnlab.github.io/mochi-kids/

Quick start

# play it locally
make serve                  # or: cd public && python3 -m http.server 8000

# run all tests
make test

Or double-click start.command from Finder on macOS.

What's here

mochi-kids/
├── public/                    # the static site GH Pages serves   → public/README.md
│   ├── index.html             # single-file jsPsych v8 experiment
│   ├── manifest.json          # 80 curated trials
│   ├── stimuli/<trial>/0..2.jpg
│   ├── audio/                 # 3 gTTS prompts (welcome / how_to_play / reminder)
│   └── images/zorpie/         # mascot GIFs (from vislearnlab/museumkiosk)
├── server/                    # optional Express + MongoDB save layer  → server/README.md
├── rendering/                 # rotation-animation pipeline (planned)  → rendering/README.md
├── tests/                     # asset-integrity + Playwright e2e       → tests/README.md
├── docs/                      # research artifacts (figures, sims)     → docs/README.md
├── .github/workflows/         # CI (test) + GH Pages (pages)           → .github/README.md
├── TESTING.md                 # full testing strategy
├── start.command              # double-click → serves locally + opens browser
└── push_to_vislearnlab.command  # one-click: gh repo create + push + enable Pages

Trial set (80 trials, gray-render only)

tier	n	dataset	content	mean adult acc
training	12	synthesized	same image × 2 + 1 different image (pop-out)	trivial
warmup	12	shapenet	easiest chair / lamp / bench	1.00
familiar	28	shapenet	8 categories: chair, lamp, bench, telephone (4 each) + car, airplane, sofa, table (3 each)	0.97
novel	28	shapegen	random sample from abstract4/abstract3/abstract2	0.91

Familiar and novel each split into 2 sub-blocks at runtime; the four sub-blocks play in random order after warmup. Each block opens with a short Zorpie intro screen.

Each manifest entry preserves human_avg_adult and rt_avg_adult so calibration analyses can use them directly. See public/stimuli/README.md for the curate logic and re-run instructions.

Audio + reward design

Spoken voice fires at three structured moments only — never per-trial, so the soundscape stays calm:

when	what plays
Consent screen	welcome.mp3
How-to-play	how_to_play.mp3
Every 10 trials (?reminder_every=N)	reminder.mp3

Reward audio = chime only. A C-major arpeggio synthesized live via Web Audio API on every correct answer. No sound on wrong (no harsh buzzer). 16-particle multicolor sparkle burst from the correct card. HUD score pill ticks up with a small pop animation. End screen shows Zorpie + a count of correct answers (no star rating).

The voice files were generated with gTTS — they're functional but robotic. Drop in real recordings at the same filenames in public/audio/ to upgrade with no code changes.

URL parameters

param	default	what it does
participantID	random kid_xxxxxxxx	Prolific / SONA / lab ID
study	mochi_kids_v1	Study tag stored with the record
save	true (auto-false on *.github.io/*.web.app/etc.)	Set to false to skip POST
submit_url	/submit	Override server endpoint
reminder_every	10	Trials between spoken reminders
break_every	20	Trials between break screens

Example: https://vislearnlab.github.io/mochi-kids/?participantID=pilot01

Data shape

Each completed session POSTs (or the kid downloads) one JSON document:

{
  "participantID": "kid_xxxx",
  "study": "mochi_kids_v1",
  "consent": { "age": "6", "agreed": true },
  "n_trials": 80, "n_correct": 64, "mean_rt": 3145.2,
  "trials": [{ "task": "mochi_oddity", "trial_id": "shapenet1234",
               "tier": "familiar", "condition": "chair",
               "correct": true, "rt": 2810.4,
               "oddity_index_orig": 1, "chosen_orig_index": 1,
               "display_order": [2, 0, 1], ... }]
}

When running via GitHub Pages (no server), the user gets a "Download my data" button on the end screen. When running with the Express server (cd server && npm start), the same payload is upserted to MongoDB on participantID.

Testing

CI runs the full suite on every push to main. Three layers:

1. Static — JS syntax (node --check), Python compile, manifest schema + image existence checks
2. End-to-end — Playwright drives a real headless Chromium through the entire experiment (consent, all 80 trials, breaks, reminders, end screen) and asserts no console errors, all RTs captured, double-clicks ignored
3. Server (optional, planned) — supertest against /submit with in-memory MongoDB

Locally: make test or bash tests/run_all.sh.

Full strategy: TESTING.md.

Deploying

• GitHub Pages (no server) — git push to main. The pages.yml workflow publishes public/ automatically. Live URL: https://vislearnlab.github.io/mochi-kids/.
• Lab server (with MongoDB save) — see server/README.md.

The static client auto-disables /submit POSTs on *.github.io, *.web.app, and similar hosts so it works offline / read-only on Pages.

Roadmap

• Static play-through w/ jsPsych v8, kid-friendly UX
• Curated 80-trial easy-tail set (familiar real objects + novel abstract shapes)
• Consent + age picker, scoped audio (welcome, how-to-play, every 10-trial reminder)
• CI + automated tests, GH Pages deploy
• Real pilot — N≈30 kids per age (4, 5, 6) + matched adult sample
• Rotation-animation manipulation (within-subjects ±45° yaw on a random half of trials — waiting on ShapeNet GLB access and shapegen meshes from Bonnen; see rendering/README.md)
• Pre-registration of the human-model crossover hypothesis (kids beat models on familiar real objects, models beat kids on novel abstracts — see docs/simulated_results.png for the predicted pattern)
• Replace gTTS prompts with a real recorded voice

Citation

If you publish work using this code or stimuli, please cite the MOCHI benchmark:

Bonnen, T., Fu, S., Bai, Y., O'Connell, T., Friedman, Y., Kanwisher, N.,
Tenenbaum, J. B., & Efros, A. A. (2024). Evaluating Multiview Object
Consistency in Humans and Image Models. NeurIPS Datasets & Benchmarks.
arXiv:2409.05862.

Acknowledgments

Mascot art (Zorpie) from brialorelle/museumkiosk. Stimuli from tzler/MOCHI on Hugging Face. Schoolbell font from Google Fonts. jsPsych v8.