Short answer
License existing footage when you need volume, real operating conditions and many sites for pretraining or task understanding, and can accept the viewpoints, labels and consent terms it was recorded under. Commission recording when you need a specific viewpoint, calibrated sensors, visible hands, planned task coverage, failure cases or consent written for AI training. Many programs combine them: archives for breadth, commissioned capture for the robot's target tasks.
The decision this page settles
The general choice between existing assets, new collection and specialist partners, across every data type, is covered in existing assets vs new collection vs specialist data. This page narrows it to one asset: video of people doing physical work. Should you license what a business already recorded, or pay to record it again on your terms?
The answer depends less on price than on which properties your training use needs, and whether those properties can be added after the fact.
What each route gives you
| Factor | Existing footage | Commissioned recording |
|---|---|---|
| Viewpoint and sensors | Whatever the original purpose needed: overhead, presenter-framed, occasionally head-mounted | Chosen to match the robot: eye-level head camera, wrist camera, multi-view, depth, IMU |
| Calibration and clock | Rarely documented | Specified: intrinsics, extrinsics, shared timestamps |
| Task coverage | What the business happened to record, mostly routine successful work | Planned against a task list, including variations, recoveries and failures |
| Labels and language | Titles, sometimes procedure IDs; seldom step boundaries | Designed in: narration, step markers, task and outcome IDs |
| Consent and notice | Given for training, quality or security purposes | Written for the intended AI use from the start |
| Realism | Real pace, interruptions and sites | Real on live work; less so when staged |
| Volume and site diversity | Potentially large, many sites, many years | Bounded by budget, recruitment and site access |
| Time to first data | After rights review and selection | After scoping, recruitment, equipment and a pilot |
Properties you can add later, and properties you cannot
This is the test that decides most cases.
| Property | Can it be added to existing footage? | Notes |
|---|---|---|
| Task and step labels | Yes | Annotation cost scales with hours; see annotating existing footage |
| Language descriptions | Yes, after the fact | Post-hoc captions are weaker than narration recorded during the work |
| Object masks and tracks | Where objects are visible | Occlusion limits coverage |
| 2D hand keypoints | Partly | Only where hands are visible and ungloved |
| Redaction | Yes | It lowers training value and does not settle permission questions |
| Camera viewpoint | No | The archive's angle is fixed |
| Camera calibration | Rarely | Only if the same cameras are still installed and can be calibrated now |
| Depth, IMU, gaze | No | Never recorded |
| Multi-sensor synchronization | No | Needs a shared clock at capture time |
| Planned failures and rare events | No | You can only search for what happens to exist |
| Consent scope for AI training | Sometimes | Re-notice or re-consent may be impractical, especially for former staff |
If a property you need sits in a "No" row, no amount of preparation will make the archive provide it. That is the clearest signal to commission.
What designed capture looks like in published datasets
Purpose-built collections show what commissioning buys. Ego-Exo4D recorded skilled activities simultaneously from egocentric and multiple exocentric cameras, with multichannel audio, eye gaze, 3D point clouds, camera poses, IMU and expert commentary (Grauman et al., 2023). Assembly101 used 8 static and 4 egocentric cameras on unscripted assembly that included mistakes (Sener et al., 2022).
Language and failures can also be designed in. EPIC-KITCHENS-100 used head-mounted cameras and "pause-and-talk" narration as its first annotation stage (Damen et al., 2020). CaptainCook4D recorded normal and intentionally erroneous recipe executions under IRB approval with written consent (Peddi et al., 2023).
Designed capture still costs effort after recording. Ego4D worked with consenting participants and de-identification procedures, and reports more than 250,000 hours of annotator effort (Grauman et al., 2021).
Existing video wins on scale. HowTo100M assembled 136 million clips from 1.22 million narrated web videos and used the automatically transcribed narration with no manual annotation (Miech et al., 2019); later work found that much of that narration does not describe what is on screen at the time. Business archives often behave similarly: abundant, weakly labeled and loosely aligned.
Where existing footage usually comes from
- Training and onboarding libraries, the most plausible existing asset; evaluate them with the corporate training video archive checklist.
- Vocational and certification recordings, which carry proficiency judgments; see vocational training and certification recordings.
- Technician body-camera and remote-assistance sessions, recorded on customer premises; see technician body-camera footage.
- Fixed process-monitoring and CCTV cameras, which rarely show hand detail; see the limits of fixed-camera footage.
What commissioning involves
A commissioned recording needs a task list, a capture specification, demonstrators who consent, site access, equipment, calibration and in-collection QA. Two early choices shape it: whether to record live operations or staged sessions (staged vs live operation recording), and how small a first run should be (pilot collection scoping).
The SourceX site describes a separate physical-world path for new recordings of hands-on work coordinated with businesses, where each opportunity has its own requirements.
A decision procedure
- Name the training use: visual pretraining, latent-action pretraining, co-training with tracked hands, step recognition, or evaluation.
- List must-have properties for that use, marking any that appear in the "No" rows above.
- Request a sample from the archive and audit it against the list.
- Commission for any use whose must-have properties fail and cannot be added.
- License and budget preparation for uses the archive passes.
- Combine where it pays: an archive for breadth, a commissioned set for target tasks and for a held-out evaluation set.
Illustrative: two requests for one task family
# Illustrative only — two ways to source data for shelf restocking
archive_license:
asset: existing training and quality-audit video, retailer_009, 2021-2025
use: visual pretraining; step recognition
must_have: hands visible in sampled frames above a buyer-set share; procedure IDs; constant frame rate
preparation: selection, step segmentation, face and screen redaction
rights: notice covered internal training; AI-training permission to be confirmed per video
commissioned_recording:
sites: store_A, store_B (live operation, off-peak)
demonstrators: consenting staff with a range of experience (operator_A01 ... operator_A12)
capture: eye-level head camera + wrist camera, calibrated, shared clock
coverage: SKU list incl. fragile and heavy items; planned recoveries
labels: narration during capture; step markers; outcome flag per attempt
rights: consent written for AI training; license defines redistribution limits
Hidden costs on each side
Existing footage: most hours are irrelevant and must be screened out; rights must be checked video by video; vendor-produced content may not belong to the business; annotation is needed before training; redaction removes useful pixels.
Commissioned recording: recruiting and briefing demonstrators; disruption to live work; equipment and calibration; consent administration; QA and re-shoots; repeat visits to cover variation.
Neither route has a universal minimum volume. What you need depends on architecture, task, embodiment and intended use.
Permissions and provenance
Existing footage was recorded under a notice that described some purpose. Ask what that purpose was, whether licensing for AI training falls within it, and who appears on camera: employees, contractors, customers or members of the public.
Commissioned recording lets consent be written for the intended use, but it still needs site access and the agreement of everyone recorded. Worker consent and identifiable people covers the diligence questions. This is not legal advice.
Questions to ask suppliers
- Why was this footage recorded, and what notice did the people in it receive?
- Which cameras, mounts and frame rates were used, and are any of those cameras still installed?
- Can clips be linked to task, procedure or work-order IDs?
- What share of the archive shows hands and handled objects up close?
- Are failed, interrupted or corrected attempts included, or edited out?
- Who produced the footage: in-house staff or an outside vendor?
- For new recording: which sites could host it, during which hours, and with how many willing demonstrators?
- For new recording: can consent be written to cover model training, licensing and retention?
Discuss your requirements
If you have worked out which properties your use needs, send SourceX the task list and must-have properties. SourceX looks for US businesses that hold suitable footage, checks the data and the supplier's licensing permissions, and coordinates the license and delivery. If no existing asset fits, SourceX can scope a commissioned collection of new recordings. SourceX reviews each request and follows up about potential sources; a request does not guarantee a match.
For background, see what human demonstration video can and cannot do for robot learning, the workplace task demonstration hub and the humanoid training data guide. The catalog page on licensing video recordings covers record-level terms.
Describe the workplace demonstration data your robotics project needs
Sources
- Ego4D: Around the World in 3,000 Hours of Egocentric Video — consent, de-identification and annotation effort in designed collection.
- Ego-Exo4D: Understanding Skilled Human Activity from First- and Third-Person Perspectives — synchronized multi-view, multi-sensor capture.
- Rescaling Egocentric Vision — narration designed into capture.
- CaptainCook4D: A Dataset for Understanding Errors in Procedural Activities — planned erroneous executions with written consent.
- Assembly101: A Large-Scale Multi-View Video Dataset for Understanding Procedural Activities — static and egocentric cameras on unscripted work with mistakes.
- HowTo100M: Learning a Text-Video Embedding by Watching Hundred Million Narrated Video Clips — existing narrated video at scale and its weak alignment.