Human Demonstration Data for Robotics: How Labs Source Egocentric Video at Scale

Human demonstration data is video of real people doing real physical tasks, recorded close enough to the body to show exactly how a hand reaches, grips, and adjusts. robomimic's research on offline learning found that demonstration quality and diversity matter as much as the algorithm, which is why labs now run sourcing operations to collect footage at scale.

Key numbers

• Claru: 10,000+ collectors across 100+ cities, and 500,000+ egocentric videos captured for a world-modeling lab (Claru)
• Perle: 49,000 vetted experts across 70 countries (Perle)
• Micro1: roughly 4,000 contributors across 71 countries producing more than 160,000 hours of video a month (RemoWork, 2026)
• Atlas Capture: 10,000+ active contributors and more than 2 million episodes recorded, with $10M+ paid out to contributors (Atlas Capture)
• Our AI Trainers program: trainers are paid $6 to $15+ per approved hour depending on the task, with the rate shown before they accept (AI Trainers page)

What counts as human demonstration data

A demonstration is a recorded instance of a task done the way a person actually does it: loading a dishwasher, folding laundry standing up, driving a familiar route, restocking a shelf. For robotics, the camera usually needs to sit where the robot's "eyes" effectively will, which is why most of this footage is egocentric (first-person), shot from a head mount, chest mount, or a camera clipped near where hands do the work.

This matters because robot policies trained on demonstrations only generalize as well as the demonstrations let them. robomimic's finding that demonstration quality and diversity strongly affect offline policy learning is the reason a lab cannot simply scrape internet video and call it training data: footage shot from the wrong vantage point, with hands out of frame, or recorded by the same person doing the same motion over and over, teaches a narrower and more brittle policy than footage collected deliberately across many bodies, kitchens, and ways of gripping a cup.

Some of that need is met by public benchmarks. Ego4D, released by Meta AI with a university consortium, is the best-known large public egocentric dataset, and it is commonly used to pretrain general egocentric perception before a lab fine-tunes on footage built around its own task list. But a public dataset was never built around any one lab's robot, its gripper, or its specific list of household and work tasks, which is where custom collection takes over.

How labs source it at scale

Once a lab knows which tasks its policy needs to see, it has to decide how to get footage of those tasks, at volume, from people who actually do them, and it has four broad ways to do that: pull from a public dataset, open a crowd app to anyone with a phone, brief a collector network that has already built a vetted contributor pool, or hand the brief to a recruiter who builds and manages the workforce directly.

The four models are not a strict ladder. A lab often runs more than one at once, pulling general pretraining footage from a public dataset while briefing a collector network or a managed workforce for the specific task set its current policy needs.

What this costs, by model

Dollar figures on the contributor side are public across most of these platforms and give a useful read on where the effort and vetting sit, even though what a lab itself pays for a given engagement is rarely published.

A few things fall out of that table. First, egocentric demonstration recording and general AI evaluation work are priced very differently: Mindrift, for comparison, advertises $15 to $100+ an hour for domain-expert evaluation work done mostly by people with advanced degrees, which is a different labor market from recording a kitchen task. Second, RemoWork's 2026 roundup found effective rates across several first-person video platforms clustering in roughly the $3 to $9 per approved hour range, with wide variation by task and vetting depth. Third, pay structure tends to track vetting depth: per-clip crowd pricing with tiers suggests lighter per-contributor screening than a model where every profile is reviewed before a brief goes out.

Quality and consent controls buyers should demand

Price tells only part of the story. A lab evaluating any of these four models should get clear answers on four points before footage starts coming in.

Brief adherence. Does the contributor see exactly what to record, how long the clip should run, and in what environment, before they start filming, or only find out after submission? A brief that specifies the task, the length, and the setting up front gives a lab a chance to catch mismatches before paying for footage it cannot use.

Approval standard. What actually counts as acceptable footage: is it stated in writing, or discovered through rejection after the fact? We define an approved hour as footage that follows the brief: the task is real, the camera is steady and well lit, the whole task is visible, and there is no idle time padding it out, and we tell trainers exactly what was accepted and why. That is the kind of standard worth asking any provider to put in writing.

Consent of people in frame. Everyone who appears in a clip needs to have agreed to it, and filming needs to be legal in the location it happens. This becomes more complicated with household and caregiving footage, where family members, children, or coworkers can end up on camera without having agreed to anything. RemoWork's 2026 review of Atlas Capture also flagged audit-based account suspensions near payout as the platform's biggest contributor risk, which is a signal worth asking any crowd-sourced provider about directly: how review and suspension actually work, and at what point in the payment cycle.

License terms. Is the footage licensed to the lab for training only, or could it end up reused or published elsewhere? A clear answer here protects both the lab's competitive position (most of this work sits under NDA) and the people being recorded. Our terms state that footage is licensed to the lab for training its models, not posted anywhere public, that faces are optional on most projects, and that a trainer can withdraw and have their profile deleted at any time.

Guides like our breakdown of what is egocentric video and why AI labs pay for it and the full explanation of what counts as an approved hour of training footage go deeper on both of these questions if a program is new to your team.

When each model fits, and when it does not

Public datasets fit early-stage research, general pretraining, and situations with no dedicated collection budget. They do not fit once a lab needs footage of a specific task, a specific environment, or a specific population the policy will actually operate around, because no public release was built for that.

Crowd apps fit broad, well-defined task categories where a lab wants high volume fast and can tolerate variable contributor quality in exchange for scale and a tiered per-clip price list. They fit less well when a lab needs guaranteed demographic or task specificity, or cannot absorb the churn and rejection rates that come with an open, self-serve pool.

Collector networks fit a lab that needs real scale with some baseline vetting across many cities or countries in a short window, and is comfortable working inside a marketplace relationship rather than a single-recruiter relationship. They fit less well when a task is narrow or specialist (a particular trade, a caregiving routine, a piece of equipment that few contributors own), because the depth of vetting applied to that narrow slice of the pool is harder to verify from outside.

A managed recruited workforce fits a lab that wants a human recruiter standing between the brief and the camera: profiles reviewed individually, tasks matched to people who actually do them day to day, consent and legal-filming status confirmed before anyone records, and in some cases a camera shipped and retrieved rather than relying on whatever phone a contributor already owns. It fits less well for a lab that simply needs a few hundred generic clips fast and has no particular interest in bespoke vetting per contributor.

Our comparison of egocentric video data providers across crowd apps, collector networks and managed workforces walks through this same trade-off in more detail, and the direct comparison between Claru and our AI Trainers program looks at one collector network against one managed workforce specifically.

A sourcing checklist for buyers

Our homepage Industries card for AI training describes the same shape in one line: a lab brief comes to us, and we recruit, vet and pay AI Trainers who record everyday tasks on a phone or a camera we ship, supplying human demonstration data to the largest AI and robotics labs in the world. The AI Trainers sign-up page shows the contributor side of that same process: sign up, get matched with a brief that states the task, the length and the hourly rate, then get paid per approved hour once footage clears review.

The short version

Human demonstration data is first-person footage of real tasks, and robotics policies need it in volume and in variety, a point robomimic's research on offline learning backs directly. Four sourcing models cover the market: public datasets like Ego4D for broad pretraining, crowd apps like Atlas Capture for fast open-sign-up scale, collector networks like Claru, Perle and Micro1 for vetted pools across many countries, and a managed recruited workforce like ours for brief-matched, person-reviewed, pay-per-approved-hour collection. Before choosing any of them, get a straight answer on brief adherence, the approval standard, consent of everyone in frame, and the license terms the footage is collected under.