Egocentric Data for Robotics: What It Is, Who Buys It, and What It Pays

Language models learned from text on the internet. Robots need something the internet has very little of: video of human hands doing real tasks, filmed from the worker's own point of view. That footage is called egocentric data, and robotics companies are now paying to collect it.

What is egocentric data?

Egocentric data is video, and sometimes motion and depth data, captured from a person's own point of view, usually with a head-mounted or chest-mounted camera or smart glasses. Because the camera sees what the worker sees, the footage shows hands, tools, and objects up close, exactly the view a robot needs to learn manipulation tasks. Researchers have used it for years; what is new is that robotics companies now buy it commercially.

This guide covers who is buying, what it pays, what buyers need, and an honest view of the opportunity for Canadian businesses, as of October 10, 2026.

Who is buying egocentric video

• NVIDIA says its GR00T N1 humanoid robot model was trained on a mix that includes egocentric human video, alongside robot and synthetic data.
• Figure announced a partnership with Brookfield to capture human video across Brookfield's homes, offices, and logistics spaces, to train its robot model.
• Tesla moved Optimus training toward employees wearing helmet-and-backpack camera rigs while doing tasks like folding shirts, according to reporting in 2025.
• Skild AI says its model can learn new skills largely by watching video, with less than an hour of robot data.

Source: NVIDIA Research, GR00T N1

Source: Figure, partnership with Brookfield

Source: eWeek, Tesla Optimus training

Source: Skild AI, learning by watching

Notice that Tesla and Figure collect much of their own footage. The open market runs mostly through data vendors and gig platforms. MIT Technology Review has described a growing gig economy of people paid to film themselves doing chores, and DoorDash launched an app in 2026 that pays couriers to film everyday activities, such as washing dishes, to train AI.

Source: MIT Technology Review, the humanoid data gig economy

Source: TechCrunch, DoorDash Tasks

What it pays, and why prices are falling

• Collection pricing: Indian collection startups price data at around US$1 to $10 per hour, depending on the task, as reported by Inc42.
• Sold to robotics labs: once packaged, footage sells for about US$15 to $50 per hour, with raw footage at the low end and fully annotated data at the top.
• Worker pay: as little as US$1 to $4 an hour in India, versus up to about US$30 an hour for experienced contributors on North American job listings.

Source: Inc42, India trains the world's robots

Source: TechCrunch, worker pay in India

Source: OpenTrain, egocentric video contributor listing

That gap is the whole story. Large collection operations in lower-cost markets have made everyday chore footage, like folding laundry and loading dishwashers, a commodity. A Canadian business cannot win on price filming the same tasks.

What buyers typically expect

Requirements vary by buyer, but collection listings and published datasets point to a common baseline:

• Video: at least 1080p at 30 frames per second, with a wide field of view and the hands clearly in frame
• Rig: a phone ultrawide lens on a head strap is a common minimum; higher-end rigs add depth sensing and motion data
• Labels: each clip broken into task steps, often with a spoken or written narration of what the worker is doing
• Metadata: scene, location type, lighting, and the task being performed
• Consent and blurring: releases from everyone filmed, with faces, screens, and documents blurred

Source: arXiv, egocentric capture with a head-mounted phone

Public research datasets show the quality bar. Meta's Ego4D includes 3,670 hours from 931 camera wearers and is released under a licence agreement that limits how it can be used. Apple's EgoDex captured 829 hours of detailed hand tracking.

Source: Ego4D paper

Source: EgoDex paper

Could trades wear cameras as part of the job?

Here is the idea we find most interesting for Canada. Instead of hiring people to film, a trades, construction, or mining business equips its crews with helmet cameras that pay for themselves on the job, and the footage becomes a licensable asset as a by-product.

There is precedent. OpenSpace, a construction technology company, has workers mount a 360-degree camera on their hard hat during site walks, and its software turns the footage into a searchable visual record of the site, used for accountability, RFIs, and inspections.

Source: Bisnow, OpenSpace hard hat cameras

That points to the right order of value. A contractor would wear a camera first for reasons that help the business today: documenting work for warranty and insurance claims, training apprentices, and getting remote help from a senior tradesperson. Licensing the footage to robotics companies is upside on top, not the reason to start.

The hurdles are real, and worth naming before anyone buys hardware:

• Worker privacy: filming employees falls under workplace privacy rules such as BC PIPA. Expect to need a written policy, clear notice, limits on retention, and real consent, plus agreement from any union.
• Customer privacy: plumbers and electricians work inside people's homes. Homeowners, children, mail, and screens all need consent or blurring.
• Mining sites: mining is a natural fit for Canada, and crews already wear helmet-mounted gear. But underground and other hazardous areas can require electronics certified for those environments, and mine sites have strict rules on devices and recording. Any camera program has to go through the site's safety and operations teams first.
• Safety equipment: hard hats are certified safety gear. Check the manufacturer's written guidance on accessories and confirm with your safety committee or WorkSafeBC before mounting anything.
• Employment status: if you pay people specifically to record, they may legally be employees. BC's employment standards guidance notes that calling someone an independent contractor does not decide the issue. Get advice before you structure payments.

Source: Government of BC, gig worker rules

A cautious way to test it

If you run a trades or operations business and want to explore this, here is the low-risk path:

• Ask first: talk to robotics data vendors about what skilled-task footage they would buy and at what price, before recording anything.
• Get advice: have a privacy and employment lawyer review your consent forms and policy.
• Record a sample: 10 to 20 hours of one task type, with consenting staff, in a controlled setting, to the spec above.
• Scale only with a commitment: move to a larger paid pilot only with a written commitment from a buyer.

Signal & Form helps Canadian businesses evaluate, package, and broker data like this. See our AI data licensing service, our breakdown of what AI companies pay for training data, and our guide to selling data to AI companies from Canada.