Choosing a delivery model

Three delivery models dominate robotics data work today: dedicated teams, crowdsourced operator networks, and hybrid configurations that combine both. The choice between them is the single highest-leverage decision in scoping a program. It determines unit cost, IP isolation strength, ramp time, and quality consistency — all four at once.

This guide is the framework we walk customers through when they’re trying to decide. It’s vendor-neutral: there are situations where each model is right, and saying “hybrid for everyone” is a sales answer, not a useful answer.

Dedicated teams

A dedicated team is operators on payroll (your data partner’s payroll, working full-time on your program). Same people every day, same rigs, same physical pod. The relationship looks more like a contracted engineering team than a marketplace.

What it’s good for: high-IP work where competitive secrecy matters; multi-year programs where operator context compounds; novel robotics platforms where operator training takes weeks; regulated work (surgical, defense) where audit trail per operator is contractual.

What it costs: typically $40–60 per operator-hour. A 10-person dedicated team running at 8 hours/day for a month: roughly $80K–$120K all-in. The unit-cost premium over crowdsource is 2.5–4x.

What you give up: ramp speed (4–6 weeks to train a dedicated team) and demographic diversity (10 operators have less variation than 200). For tasks that benefit from population-level diversity, this matters.

Crowdsource

Crowdsource is a vetted operator network with hundreds to thousands of operators, drawn into a program based on tier qualification and availability. Different operators work on the program week to week. Same NDA framework as dedicated; different operational model.

What it’s good for: programs that need population diversity (different hand sizes, different cultural backgrounds for language tasks, different demographic coverage); high-volume programs where unit cost dominates; programs where the task is standardized enough that operator-specific context doesn’t compound.

What it costs: typically $8–16 per task-equivalent (varies wildly by data class). For high-volume programs, unit costs can be 2–4x lower than dedicated.

What you give up: IP isolation guarantees (no individual operator works exclusively on your program); deep program context (no operator has 6 months of program-specific learnings); consistency of quality across operators (calibration is the bar, but variance is higher).

Hybrid

Hybrid is dedicated spine plus crowdsource edge. The spine is a small dedicated team (typically 3–10 operators) who hold the IP-critical work, the calibration sets, and the QA function. The edge is crowdsource operators handling volume against the spine team’s acceptance criteria.

Most production programs end up here. The reason: pure dedicated is too expensive at scale, pure crowdsource has too much variance for production-grade output, and hybrid keeps the best of both.

What it’s good for: production programs that need to scale beyond 10 dedicated operators worth of throughput; programs with sensitive IP that still need cost-effective scale; programs where some data classes (high-IP) need dedicated and other data classes (high-volume edge cases) work fine on crowdsource.

What it costs: the math depends on the spine/edge split. An 80/20 split (80% volume on crowdsource, 20% on dedicated spine) typically lands at a blended unit cost roughly 1.4–1.8x crowdsource pricing — still cheaper than pure dedicated.

What you give up: operational complexity. Hybrid programs have two operator pools, two QA standards (the spine catches what the edge misses), and more program-management overhead. If your team can’t handle the coordination, hybrid becomes worse than either pure option.

The decision framework

Three questions, in order:

1. Is competitive IP isolation a contractual requirement? If yes (surgical, defense, frontier-AI), start with dedicated. Don’t worry about cost yet.

2. If IP isolation is not strict, does your volume target exceed what a 10-person dedicated team can produce? If yes, you’ll need either hybrid or pure crowdsource. If volume target is moderate (e.g., 5,000 trajectories/month), dedicated may still fit.

3. If you need volume and don’t need strict IP isolation, does your task benefit from population diversity? Pick crowdsource. If your task is consistency-bound and population diversity doesn’t help (most humanoid teleop), pick hybrid.

The end-state for most production programs is hybrid. But the path to get there often starts with a dedicated pilot to calibrate, then hybrid to scale. Going direct to crowdsource is rare and usually means the task is well-understood (annotation, evaluation) rather than novel (teleop, demonstration).

When you’re ready to scope, tell us the task, the volume, and the IP requirements. We come back with a model recommendation and pricing within one business day.