Turn idle moments into useful physical-world training data. Contributors capture everyday objects and environments with their phone; the app labels what they shot, packages it for buyers, and pays them when it passes review.
| Capture | Review | Data package | You earned |
|---|---|---|---|
 |
 |
 |
 |
- Capture — Open the app and record a short walk-around clip of anything nearby (a bike rack, storefront, street sign, furniture). No task list required.
- Auto-label — On submit, a vision model reads the video frame and location metadata, then assigns a category and short subject label. Contributors never type anything.
- Review — An operator checks quality (angles, lighting, duplicates, supported subjects) and accepts, rejects, or asks for a retry with plain-language feedback.
- Earn — Accepted captures accrue rewards. Totals and per-submission history live on the Earnings tab; payout rails are planned for a later release.
- Package — After submit, the app previews a buyer-ready data package: original media, key frames, robot-vision labels, quality scores, and an export manifest.
Optional Tasks boost earnings when buyers need specific categories or locations. Progress adds streaks, levels, badges, and challenges so capture feels rewarding without turning into technical work.
Robotics, physical AI, simulation, and spatial-mapping teams need large volumes of real-world object and environment data. Sending centralized field teams everywhere is slow and expensive. Millions of people already carry capable cameras and have idle windows — waiting at a bus stop, walking a neighborhood — but there is no simple, trusted way to capture useful data and get paid for it.
OG Data connects that demand with distributed contributors: easy on the phone, structured on the buyer side.
| Contributors | Data buyers | |
|---|---|---|
| Who | Everyday people with smartphones | Robotics / physical-AI companies, simulation teams, research ops |
| Goal | Capture something nearby, submit, see feedback, watch earnings grow | Get targeted, high-quality physical-world captures at scale without a field team |
| Experience | Camera-first, game-like, no jargon | Campaigns, quality signals, packaged datasets (operator-facing today) |
Contributor UX comes first. Buyer complexity stays behind operator tools until the capture loop is solid.
Labeling is automatic at capture time, not a separate manual step for contributors.
- Input: One frame from the recorded clip plus GPS coordinates (when available).
- Model: OpenAI vision (
gpt-4o-miniby default) classifies into buyer-meaningful buckets:vehicle,furniture,storefront,signage,nature,infrastructure,indoor_object,other. - Output: A category and a short subject label (e.g. “red fire hydrant”) attached to the submission metadata.
- Fallback: If the API key is missing, the request times out, or parsing fails, the capture still submits as
uncategorized. Reviewers always see the raw media.
This is MVP labeling — enough to sort captures and demo buyer value. Heavier annotation, 3D reconstruction, and production review automation are explicitly out of scope unless the PRD changes.
| State | Meaning |
|---|---|
| Pending review | Submitted, waiting for operator decision |
| Accepted | Quality approved; reward assigned (amount can vary by task complexity and demand) |
| Needs retry | Fixable issue — retake with the feedback shown |
| Rejected | Does not meet guidelines |
The Earnings tab shows total earned, pending balance, and reward history per submission. Payout integration (cash-out, thresholds, compliance) is Post-MVP; rewards accumulate in the app until then.
Suggested Tasks display an expected reward range up front so contributors know what a campaign is worth before they start.
After submit, the data package screen shows what a buyer would receive — without claiming production-grade 3D reconstruction:
- Original capture (video or photos)
- Extracted key frames with detection-style overlays
- Capture prompt and context metadata (location, time, category, AI label)
- Quality report (coverage, clarity, consistency — scored and graded)
- Export manifest listing media + JSON sidecars
The builder is deterministic and client-side for the demo. It previews the shape of robotics-ready training data; final export formats stay flexible per buyer agreement (see Post-MVP P3 — Dataset Packaging in the PRD).
| Tab | Purpose |
|---|---|
| Capture | Home screen — record and submit free captures |
| Tasks | Optional location-aware suggested tasks with reward ranges |
| Progress | Streaks, levels, badges, challenges, leaderboard |
| Submissions | History, status, reviewer feedback, retry entry |
| Earnings | Totals, pending rewards, payout-coming-soon notice |
Operator routes (not in the main tab bar): Review queue, Campaigns.
Expo SDK 56 · React Native · TypeScript · expo-router · expo-camera · OpenAI vision (dev) · jest-expo
npm install
npm start # press i / a / w for iOS / Android / webnpm run typecheck
npm run lint
npm testAI labeling (optional, dev): set EXPO_PUBLIC_OPENAI_API_KEY in .env or your shell so free captures get real categories instead of uncategorized. Submissions work without it.
Note: Video capture uses native camera APIs — use iOS or Android for the full capture flow; web shows a fallback message.
src/
app/ # expo-router routes (thin — delegate to features)
features/
capture/ # Free capture + guided task capture flows
submissions/ # History, status, retry
review/ # Operator review queue
rewards/ # Earnings screen
data-package/ # Post-submit buyer package preview
opportunities/ # Suggested tasks list
progression/ # Streaks, levels, badges, challenges
onboarding/ # First-run contributor intro
campaigns/ # Operator campaign management
shared/ # Types, API client, storage, UI tokens, AI helpers
testing/ # Fixtures for tests and mock API
docs/
screenshots/ # README app walkthrough (extracted from demo recording)
Scope, milestone status, and Post-MVP roadmap: README_PRD.md.