Home / How we collect / Crowdsource
Delivery model 02
Crowdsource
Distributed operator network for scene diversity, geographic spread, and fast scale-up. Per-task or per-hour pricing.
What is crowdsource collection?
Crowdsource uses our distributed network of 8,000+ vetted operators across 12 countries. Tasks are routed by skill level, scene diversity needs, and geography — giving your dataset the environmental variation a single lab can’t match.
When to choose crowdsource
- Scene diversity — kitchens, offices, retail, and homes across continents
- Fast scale-up — go from spec to first data units in 3 days
- Elastic volume — scale from 50 to 5,000 operators without hiring
- Geographic spread — lighting, layouts, and objects your lab doesn’t have
What you get
- Per-task or per-hour pricing — pay for what you use
- Skill-matched routing — operators selected by task complexity
- Automated QA — every submission scored before delivery
Best for
Teams that need maximum diversity and rapid scale — when your model needs to generalize across environments, not just perform in one lab.
8,000+ operators across 12 countries.
3 days from spec to first units.
Per-task or per-hour pricing.
Built for
When you need scale, diversity, or both — fast
Three patterns where a distributed network outperforms a single dedicated studio.
Long-tail and edge cases
Failure modes, weather extremes, lighting conditions, and scenarios your dedicated team can’t reproduce in a studio.
Geographic and demographic spread
Scenes captured across countries, climates, and cultural contexts that single-studio dedicated teams can’t reach.
Volume bursts and fast scale
Pre-launch dataset expansion, model fine-tune sprints, and benchmark sweeps where the constraint is operator hours, not quality bar.
How it works
From spec to first units in three days
No recruiting cycle, no training program. Operators are already vetted and tier-qualified.
Spec the task
Task description, acceptance criteria, per-unit payment, and qualification filters.
Recruit from pool
Qualified operators selected by domain, hardware access, language, and geography.
Calibration batch
Small sample reviewed by your team. Criteria lock before bulk collection begins.
Data flowing
Continuous task assignment. Live throughput dashboard. On-demand QA samples.
Network composition
A distributed network, not a marketplace
Operators are pre-vetted, NDA-signed, and tier-qualified. No race to the bottom.
Vetted operators
Pre-qualified across domains
Countries
Geographic and time-zone coverage
Languages
Native-fluency operators per language
Qualification tiers
Associate → Senior → Lead → Specialist
Network ops manager
Your single point of contact
Operators progress between tiers based on accuracy across 500+ baseline tasks. Specialist tier reserved for surgical, AV, and ITAR-cleared work.
Quality control
Four pillars hold the floor
Distributed doesn’t mean uneven. Every layer of the network is measured and adjusted continuously.
Qualification gates
Operators progress through tiers by hitting accuracy thresholds across 500+ baseline tasks. No tier jumping for a hot project.
Calibration batch
Every program starts with a 200-task batch reviewed jointly. Acceptance criteria lock before bulk collection begins.
Continuous spot QA
Random 5% sample re-reviewed by senior operators. Below-threshold operators are paused, retrained, or rotated out.
Throughput dashboard
Live unit counts, per-operator accuracy, geographic distribution. Anomaly alerts pushed to your team.
Compare delivery models
When to choose crowdsource
Pick crowdsource when the constraint is operator hours, geographies, or task variety.
| Criteria | Dedicated | CrowdsourceRECOMMENDED FOR SCALE + DIVERSITY | Hybrid |
|---|---|---|---|
| IP control | Full exclusivity. Operators NDA’d. | Per-batch NDA. | Mixed — dedicated core, crowdsourced edge. |
| Operator quality | Trained on your spec to a bar you set. | Tier-qualified. Variable. Aggregated. | Best for spine + scale for diversity. |
| Ramp time | 4 weeks SOW to production. | 3 days spec to first units. | 2–3 weeks. Dedicated first, crowd added. |
| Hardware | Your rig in our studio, or our matching rig. | Operator-provided or our standard rigs. | Mix of both per task type. |
| Best for | Surgical, AV, defense, industrial. | Long-tail, scene diversity, geo spread. | Most production foundation models. |
| Pricing | Per FTE-month. | Per task or per hour. | Custom SOW. |
What our partners say
FAQ
Questions about crowdsourced data collection
Further reading
From the blog
How to Scale Teleop Data Collection Without Losing QualityManaging quality across a distributed crowdsource operator network.
From the blog
Robot Data Annotation: A Practical Guide for ML TeamsAnnotation at scale with crowdsource operator pools.
Spin up a crowdsourced collection
Tell us the task, the volume, and the spread. First units land in your bucket within four days.