How to Choose an Egocentric Data Collection Vendor (and What Drives Cost per Hour)

Last updat­ed: Octo­ber 2026

Choos­ing an ego­cen­tric data col­lec­tion ven­dor is now a core sourc­ing deci­sion for robot­ics and embod­ied AI teams. First-per­son video is cheap to record and expen­sive to get right. The wrong rig, a miss­ing cal­i­bra­tion file or a con­sent gap can make hun­dreds of hours unus­able for train­ing.

This buy­er’s guide is for robot­ics, phys­i­cal AI and com­put­er vision leads who need humanoid robot train­ing data or oth­er first-per­son video at scale. You will get a cap­ture rig com­par­i­son, the eight-point CAPTURES ven­dor check­list, a cost-dri­ver table that explains what moves ego­cen­tric video cost per hour, a pilot plan, RFQ ques­tions and the red flags we see most often.

Want to see what good looks like first? Request a sam­ple clip set straight off our rigs, with the cal­i­bra­tion files attached.

What does an egocentric data collection vendor do?

An ego­cen­tric data col­lec­tion ven­dor recruits and con­sents par­tic­i­pants, equips them with head-mount­ed or wear­able cam­eras, cap­tures first-per­son video of real tasks in real envi­ron­ments, and deliv­ers QA-checked, meta­da­ta-tagged footage to your spec. Good ven­dors also han­dle facil­i­ty per­mis­sions, bystander pri­va­cy, sen­sor cal­i­bra­tion, syn­chro­niza­tion and option­al anno­ta­tion.

In short: the ven­dor turns your cap­ture spec into accept­ed hours of train­ing data. Record­ing is the easy part. The work sits in the oper­a­tions around it.

A full-ser­vice ven­dor typ­i­cal­ly owns six jobs:

1. Recruit­ment and con­sent. Find­ing par­tic­i­pants and the busi­ness­es or house­holds where they work, then col­lect­ing writ­ten con­sent from both.
2. Device logis­tics. Ship­ping, charg­ing, track­ing, replac­ing and retriev­ing rigs across many sites.
3. Cal­i­bra­tion and sync. Log­ging cam­era intrin­sics and lens dis­tor­tion per device, and keep­ing video, audio and IMU streams on one clock.
4. Cap­ture man­age­ment. Train­ing par­tic­i­pants on your task pro­to­col and super­vis­ing ses­sions on site or remote­ly.
5. Meta­da­ta and QA. Tag­ging each file with task, envi­ron­ment and device data, then check­ing it against accep­tance cri­te­ria.
6. Deliv­ery and anno­ta­tion. Pack­ag­ing files in your for­mat, with option­al hand pose, action seg­ments or object tags.

If you are new to the for­mat, our primer on what ego­cen­tric video is and how it is used cov­ers the basics. Research datasets show what this looks like at scale. Ego4D, for exam­ple, holds 3,670 hours of dai­ly-life video from 923 par­tic­i­pants in 74 loca­tions across 9 coun­tries, accord­ing to the Ego4D project site. Pub­lic data like this is a start­ing lay­er. Our guide to the Ego4D dataset and when you need cus­tom data explains where it stops fit­ting a com­mer­cial robot.

Which capture rig should your vendor use?

Pick the cheap­est rig that still records the sig­nal your mod­el needs. Head-mount­ed smart­phones suit high-vol­ume RGB cap­ture. Action cam­eras suit harsh or low-light sites. Stereo and depth rigs suit 3D tasks. VR head­sets add rough hand track­ing. Cus­tom rigs match a robot­’s own cam­era field of view.

For head-mount­ed cam­era data col­lec­tion, the rig decides more of your bud­get and your data qual­i­ty than almost any oth­er choice. It sets res­o­lu­tion, field of view, dis­tor­tion, sen­sor streams and how long a par­tic­i­pant can com­fort­ably record. Ask every ego­cen­tric data col­lec­tion ven­dor which device fam­i­lies they run in pro­duc­tion today, not which ones they could source.

Rig	Best for	Trade-offs
Head-mount­ed smart­phone	Bulk RGB hours for house­hold and work­place tasks; usu­al­ly the low­est cost per hour	Heat and bat­tery lim­its on long ses­sions; each phone needs its own cal­i­bra­tion file
Action cam­era	Reli­a­bil­i­ty-crit­i­cal, out­door or low-light work where footage qual­i­ty dri­ves review	Wide fish­eye lens­es need dis­tor­tion cal­i­bra­tion; elec­tron­ic sta­bi­liza­tion must be switched off for geom­e­try-sen­si­tive use
Stereo or depth rig	3D scene under­stand­ing, depth-aware grasp­ing, stereo ego­cen­tric video	Base­line cal­i­bra­tion, tight frame sync, heav­ier rigs and much larg­er files
VR or mixed-real­i­ty head­set	Approx­i­mate hand track­ing with­out sep­a­rate pose anno­ta­tion	Weight and com­fort lim­it ses­sion length; raw cam­era access varies by device
AI smart glass­es	Hands-free, gaze-aligned cap­ture for pilot batch­es	Bat­tery, stor­age and clip-length lim­its; small­er fleets
Cus­tom rig matched to robot FOV	Embod­i­ment match­ing, when your robot­’s cam­era height, lens and mount­ing must be mir­rored	Build lead time, few­er units and a high­er cost per hour

Device diversity versus device consistency

There is a real ten­sion here. Ego4D delib­er­ate­ly used sev­en dif­fer­ent off-the-shelf head-mount­ed cam­eras so that mod­els would not over­fit to a sin­gle cap­ture device, per the pro­jec­t’s own FAQ. A robot­ics team fine-tun­ing one pol­i­cy for one robot often wants the oppo­site: one rig, one lens and one mount­ing height. Decide which you need before you request quotes, because it changes the fleet, the cal­i­bra­tion work­load and the price.

Sensor streams beyond RGB

An ego­cen­tric dataset for robot­ics now often car­ries more than video. Ego-Exo4D pairs its 1,286 hours of video with mul­ti­chan­nel audio, eye gaze, 3D point clouds, cam­era pos­es and IMU data, accord­ing to the Ego-Exo4D paper. Each extra stream is use­ful only if it is syn­chro­nized and cal­i­brat­ed. An IMU on its own clock is noise, not sig­nal.

How to evaluate an egocentric data collection vendor: the CAPTURES checklist

Eval­u­ate an ego­cen­tric data col­lec­tion ven­dor with eight checks: con­sent records, accep­tance cri­te­ria, a paid pilot, a time-sync spec, an unedit­ed raw sam­ple, rig cal­i­bra­tion files, enforce­able legal com­pli­ance and pri­va­cy screen­ing. A ven­dor that can­not show all eight before the pur­chase order is ask­ing you to car­ry the risk.

We built CAPTURES from the ques­tions buy­ers’ legal, ML and pro­cure­ment teams ask us most often. Use it as a score­card in ven­dor calls.

• C: Con­sent records. Ask for the actu­al par­tic­i­pant con­sent form and the facil­i­ty-own­er agree­ment. Both should name AI train­ing as the pur­pose, explain with­draw­al and cov­er bystander pro­to­cols on site.
• A: Accep­tance cri­te­ria before cap­ture. Agree in writ­ing what makes a clip accept­able: min­i­mum dura­tion, hands in frame, light­ing, blur lim­its, task com­ple­tion and meta­da­ta fields. Cri­te­ria writ­ten after record­ing invite dis­putes.
• P: Pilot of 25 to 100 hours. Large enough to expose real fail­ure modes, small enough to walk away from. It should be billed on accept­ed hours only.
• T: Time-sync spec. Get the stat­ed tol­er­ance for cam­era-to-cam­era and cam­era-to-IMU sync, and how it is achieved: hard­ware trig­ger, shared clock or post-hoc align­ment.
• U: Unedit­ed raw sam­ple. Request clips straight off the rig, not a high­light reel. Raw files show dropped frames, expo­sure swings and real hand vis­i­bil­i­ty.
• R: Rig cal­i­bra­tion files. Expect intrin­sics and dis­tor­tion for every device, plus extrin­sics and base­line for stereo and cam­era-to-IMU trans­forms where an IMU is used.
• E: Enforce­able com­pli­ance. The ven­dor should map its con­sent and stor­age to the laws that apply in each cap­ture coun­try, such as Indi­a’s DPDP Act and the EU GDPR.
• S: Screen­ing for pri­va­cy. Ask how faces of bystanders, screens and doc­u­ments are blurred, who reviews the redac­tion and what the rework rule is.

What the law expects from consent

Indi­a’s Dig­i­tal Per­son­al Data Pro­tec­tion Act was enact­ed on 11 August 2023. The DPDP Rules, 2025 were noti­fied in Novem­ber 2025 with an 18-month phased com­pli­ance time­line. They require stand­alone con­sent notices that explain the spe­cif­ic pur­pose of col­lec­tion, per the Gov­ern­ment of India release of 14 Novem­ber 2025. As of Octo­ber 2026 that win­dow is still run­ning, so build to the Rules now rather than lat­er.

For cap­ture involv­ing peo­ple in the EU, the Gen­er­al Data Pro­tec­tion Reg­u­la­tion (EU) 2016/679 requires a law­ful basis for pro­cess­ing. Where that basis is con­sent, the con­troller must be able to demon­strate it (Arti­cle 7(1)) and the per­son can with­draw it at any time (Arti­cle 7(3)). In prac­tice, that means per-file con­sent records and a way to pull a par­tic­i­pan­t’s footage on request.

Want the check­list as a work­ing doc­u­ment? Ask us for the CAPTURES check­list and RFQ tem­plate, and we will send the ver­sion our team uses on scop­ing calls.

What drives egocentric video cost per hour?

Ego­cen­tric video cost per hour is dri­ven main­ly by rig type, sen­sor sync and cal­i­bra­tion needs, envi­ron­ment access, the num­ber of coun­tries and par­tic­i­pant quo­tas, the anno­ta­tion lay­er and the QA accep­tance rate. Res­o­lu­tion, task script­ing, con­sent work and deliv­ery for­mat add small­er amounts. Vol­ume spreads fixed set­up costs.

No rep­utable ego­cen­tric data col­lec­tion ven­dor can give a firm price before see­ing the full spec, and we do not pub­lish rates for that rea­son. What you can do is under­stand which levers move the num­ber, then decide which ones your mod­el tru­ly needs.

Cost dri­ver	What push­es cost up	Impact
Rig type	Stereo or depth rigs, VR head­sets, smart glass­es or a cus­tom rig matched to the robot cam­era FOV, com­pared with a smart­phone head-mount or action cam	High
Res­o­lu­tion and fps	4K or 60 fps cap­ture increas­es stor­age, upload time and review effort	Med
IMU, sync and cal­i­bra­tion	Hard­ware-synced IMU, tight mul­ti-cam­era sync and per-ses­sion cal­i­bra­tion checks	High
Envi­ron­ment access and per­mits	Fac­to­ries, farms, kitchens and retail floors need own­er sign-off, site vis­its and safe­ty induc­tions	High
Coun­tries and par­tic­i­pant diver­si­ty	Each added coun­try brings local coor­di­na­tors, legal review and demo­graph­ic quo­tas	High
Task script­ing	Step-by-step scripts, cov­er­age quo­tas per task and re-shoots for missed steps	Med
Con­sent and bystander blur­ring	Par­tic­i­pant plus facil­i­ty con­sent, sig­nage and face or screen redac­tion review	Med
Anno­ta­tion lay­er	Hand pose, action seg­ments with start and end times, object tags or nar­ra­tion	High
QA accep­tance rate	Stricter cri­te­ria mean more footage is reject­ed and absorbed in the price	High
Deliv­ery for­mat	Rec­ti­fied stereo, cus­tom con­tain­ers, transcodes and per-clip JSON meta­da­ta	Low
Vol­ume	Small batch­es car­ry the full set­up cost; larg­er pro­grams spread it across more hours	Med

Per accepted hour versus per recorded hour

A record­ed hour is any hour of footage that came off a rig. A deliv­ered hour is what remains after trim­ming and pack­ag­ing. An accept­ed hour is footage that passed your writ­ten accep­tance cri­te­ria. Only the last one trains your mod­el.

When a ven­dor prices per record­ed hour, you pay for blurred clips, missed steps and ses­sions where the hands left the frame. Your effec­tive cost is the record­ed-hour rate divid­ed by the share of footage you accept. As a sim­ple illus­tra­tion, if you accept 70 of every 100 record­ed hours, each usable hour costs about 1.4 times the quot­ed rate.

Pric­ing per accept­ed hour revers­es that. The ven­dor absorbs reject­ed footage, so the unit price may look high­er at first, but your bud­get stops mov­ing with qual­i­ty swings. It also aligns incen­tives: the ven­dor earns more only by cap­tur­ing clean footage the first time.

Where buyers overspend

• Stereo by default. If your pol­i­cy trains on monoc­u­lar RGB, stereo ego­cen­tric video adds cost and file size with­out adding sig­nal.
• Dense anno­ta­tion too ear­ly. Col­lect and accept the footage first, then anno­tate a sam­ple to see which labels change mod­el behav­ior.
• Too many coun­tries in the pilot. Prove the spec in one or two loca­tions, then add diver­si­ty once accep­tance rates are sta­ble.

How should an egocentric pilot run?

A good ego­cen­tric pilot runs in five steps: lock the spec, review an unedit­ed raw sam­ple with cal­i­bra­tion files, cap­ture 25 to 100 hours billed only on accept­ed hours, review rejects and tight­en the spec, then scale with dai­ly uploads and rolling QA. Each step has a clear exit deci­sion.

1. Spec. Define tasks, envi­ron­ments, rig, res­o­lu­tion and frame rate, sen­sor streams, meta­da­ta schema and accep­tance cri­te­ria. Include nar­ra­tion lan­guage if you need spo­ken task descrip­tions.
2. Raw sam­ple. Review unedit­ed clips from the exact rig you will use, with its cal­i­bra­tion file. Check hand vis­i­bil­i­ty, blur and expo­sure on your own tool­ing.
3. Pilot of 25 to 100 hours. Run a small batch across a few sites. Ask for per-file QA sta­tus and rea­sons for every reject­ed clip.
4. Review and fix. Sort rejects by cause. Most issues trace back to mount­ing height, light­ing or vague task steps, and each one becomes a spec change.
5. Scale. Move to dai­ly uploads with rolling QA and a week­ly report on accept­ed hours, rejects and cov­er­age by task and envi­ron­ment.

Dai­ly uploads mat­ter more than they sound. They let your ML team spot drift with­in days, not after a month of footage has piled up. Any ego­cen­tric data col­lec­tion ven­dor that is con­fi­dent in its oper­a­tions should accept these terms.

What questions should you ask in an RFQ?

A strong RFQ asks each ego­cen­tric data col­lec­tion ven­dor to prove capa­bil­i­ty in writ­ing: device fam­i­lies in pro­duc­tion, cal­i­bra­tion and sync meth­ods, con­sent tem­plates, accep­tance cri­te­ria, QA process, legal basis per coun­try, capac­i­ty per week and pric­ing unit. Ask for a raw sam­ple with every response so you com­pare footage, not promis­es.

1. Which device fam­i­lies do you run in pro­duc­tion today, and how many units of each?
2. Can you share an unedit­ed raw clip and its cal­i­bra­tion file from the rig you pro­pose?
3. How do you syn­chro­nize video, audio and IMU, and what tol­er­ance do you guar­an­tee?
4. Do you dis­able elec­tron­ic sta­bi­liza­tion and oth­er in-cam­era pro­cess­ing on request?
5. Can we see your par­tic­i­pant con­sent form and your facil­i­ty-own­er agree­ment?
6. How do you han­dle bystanders, screens and doc­u­ments in frame?
7. Which laws apply in each pro­posed cap­ture coun­try, and how does your con­sent map to them?
8. Do you price per record­ed, deliv­ered or accept­ed hour? How are accep­tance cri­te­ria agreed?
9. What week­ly capac­i­ty can you com­mit to, and what hap­pens if you miss it?
10. What meta­da­ta ships with each file, and in what for­mat?
11. Can you add anno­ta­tion lay­ers such as hand pose, action seg­ments or object tags?
12. Who owns the footage and what license do we receive for mod­el train­ing?

Red flags when choosing an egocentric data collection vendor

Walk away, or at least slow down, if you see any of these:

• Edit­ed siz­zle reels only. Pol­ished clips hide dropped frames, blur and expo­sure swings.
• No cal­i­bra­tion files. With­out intrin­sics and dis­tor­tion, geom­e­try-depen­dent mod­els can­not use the footage prop­er­ly.
• Vague sync answers. “We align it lat­er” with no stat­ed tol­er­ance usu­al­ly means mis­aligned IMU and video.
• Con­sent forms “avail­able after sign­ing”. Your legal team should review them before any con­tract.
• Pric­ing per record­ed hour with no writ­ten accep­tance cri­te­ria. You car­ry all the qual­i­ty risk.
• No bystander pro­to­col. Work­places and pub­lic spaces always con­tain peo­ple who did not sign up.
• Unlim­it­ed capac­i­ty claims. Real capac­i­ty is bound­ed by rigs, coor­di­na­tors and sites. Ask for num­bers per week.
• Unclear data rights. If the license for mod­el train­ing is not writ­ten down, assume it does not exist.

How Graveiens AI runs egocentric programs

As an ego­cen­tric data col­lec­tion ven­dor, Graveiens AI runs man­aged ego­cen­tric video data col­lec­tion built around explic­it con­sent and pay-for-approved-deliv­er­ables billing. The work­flow has four stages, and each one maps to a CAPTURES check.

1. Par­tic­i­pant onboard­ing. Par­tic­i­pants and facil­i­ty own­ers sign explic­it con­sent for AI train­ing use. Coor­di­na­tors brief them on the task pro­to­col and on-site bystander rules in their own lan­guage.
2. Cap­ture man­age­ment. We match the device to the spec from mul­ti­ple device fam­i­lies, includ­ing head-mount­ed smart­phone rigs, action cam­eras, stereo and depth rigs, VR head­sets, AI smart glass­es and cus­tom rigs built to a robot­’s cam­era field of view. Cal­i­bra­tion is logged per device and ses­sion.
3. Meta­da­ta tag­ging. Every file car­ries task, envi­ron­ment, device and ses­sion meta­da­ta to your schema, with nar­ra­tion in 25+ lan­guages where need­ed.
4. QA. Footage is checked against accep­tance cri­te­ria agreed before cap­ture. You are invoiced only for the hours you approve.

Our cap­ture net­work cov­ers indus­tri­al sites, agri­cul­tur­al work, house­holds, retail floors and kitchens. When a pro­gram needs labels, our data anno­ta­tion team adds hand pose, action seg­ments or object tags under the same QA process. Teams that also need robot-side demon­stra­tions can read how tele­op­er­at­ed demon­stra­tions com­ple­ment first-per­son video.

Graveiens AI is ISO 9001 cer­ti­fied. We do not quote prices before we see a full spec, because rig, sen­sors, sites and accep­tance cri­te­ria change the num­ber. Send the spec and we will scope it hon­est­ly, includ­ing what we would leave out.

Sources

1. Ego4D con­sor­tium, “Ego­cen­tric 4D Per­cep­tion (Ego4D)” project site and FAQ (accessed Octo­ber 2026). https://ego4d-data.org/
2. Grau­man et al., “Ego-Exo4D: Under­stand­ing Skilled Human Activ­i­ty from First- and Third-Per­son Per­spec­tives,” arXiv:2311.18259 (v4, Sep­tem­ber 2024). https://arxiv.org/abs/2311.18259
3. Press Infor­ma­tion Bureau, Gov­ern­ment of India, “Gov­ern­ment noti­fies DPDP Rules to empow­er cit­i­zens and pro­tect pri­va­cy,” 14 Novem­ber 2025. https://www.pib.gov.in/PressReleasePage.aspx?PRID=2190014
4. Reg­u­la­tion (EU) 2016/679 (Gen­er­al Data Pro­tec­tion Reg­u­la­tion), EUR-Lex. https://eur-lex.europa.eu/eli/reg/2016/679/oj

About the author

Graveiens AI Edi­to­r­i­al Team. The Graveiens AI Edi­to­r­i­al Team writes from day-to-day work run­ning con­sent-backed data pro­grams for AI and robot­ics teams, includ­ing first-per­son video cap­ture, anno­ta­tion and mul­ti­lin­gual voice data. Our guides focus on the oper­a­tional detail buy­ers need: rigs, con­sent, QA gates, file for­mats and how pric­ing is built.