Real-world human task data for physical AI
Cogmelt captures task demonstrations across home, trade, and industrial environments, then structures them into workflow-specific datasets, evaluation slices, and pilot-ready data products.
Place one hero image here
Use one image only. Save it as assets/images/home-hero.jpg and it will appear here automatically.
One platform, three structured task domains
Cogmelt is broad in supply and disciplined in packaging. The platform accepts task demonstrations from home, trade, and industrial contexts, but sells them as workflow-specific data products rather than a single undifferentiated video pool.
Laundry, cleaning, kitchen prep, organization
Laundry folding, dishwasher loading, wiping, shelf sorting, ingredient handling, and other household routines.
Maintenance, tools, residential workflows
Screened trade contributors can capture lawful maintenance actions, tool handling, fixture work, and repair-adjacent routines.
Pick / orient / place and machine-adjacent tasks
Light-industrial, workshop, and production-adjacent handling workflows for structured physical-AI collection.
Mission-driven collection, review, and payout
Cogmelt does not accept random uploads. Contributors choose a mission, follow task-specific rules, label their submission, and move through review, correction, approval, and payout states.
Mission launched
Admin defines caps, trust tier, labels, payout logic, and review checklist.
Contributor records
A contributor chooses a mission and records one clear task episode.
Labels submitted
Required fields and mission-specific labels are added before submission.
Admin review
Reviewers score framing, mission match, privacy, duplication risk, and labels.
Decision recorded
Submission is approved, rejected, or sent back with correction instructions.
Payout & packaging
Accepted clips move into payout queues and can later be grouped into export packages.
Structured outputs instead of raw footage
Cogmelt packages accepted mission outputs into dataset packs, evaluation slices, pilot collections, and failure/recovery sets. The platform stores mission snapshots, label schema versions, and review outcomes for each submission.
Accepted-submission compensation with quality controls
Contributors are paid for accepted assets, not effort hours. Approved clips can earn a base mission payout, a quality bonus, a scarcity bonus, plus delayed holdback and reserve logic managed in the admin payout queue.