Short answer
A corporate training video archive is worth licensing for robot learning when its clips show the task being done with hands and objects in view, map to named procedures, cover real variation, and come with clear ownership and on-camera consent. Such archives can support task understanding, step recognition and video-language pretraining; they never supply robot actions or forces. Pass or fail a sample against the checklist below before negotiating.
What these archives usually hold
Learning and development teams accumulate several kinds of video, and only some of it shows physical work:
- How-to and job-skills clips filmed on the floor, often the useful core.
- Onboarding modules that mix floor footage with presenters and slides.
- Safety and compliance training, often staged and short on hand detail.
- Recorded classroom sessions and webinars, which are mostly talking heads.
- Slide-based e-learning with voiceover or screen capture, which has no manipulation content.
- Vendor-produced courses and equipment-maker videos, which the business may license rather than own.
An archive described by its total hours tells you little. The runtime that shows hands working on objects is what matters, and it is often a minority.
What an archive can and cannot support
Good floor footage with narration can support visual pretraining, step recognition, video-language learning, task planning and evaluation sets for task understanding. It contains no robot joint states, actions or contact forces, its camera rarely matches a humanoid's, and it usually shows only successful routine work. The wider picture is on what human demonstration video can and cannot do for robot learning.
The checklist
Run it on a random sample drawn by you, not on clips the supplier selected. Thresholds are set by the buyer for the intended use.
Content
| Check | Pass if | Evidence to request |
|---|---|---|
| Manipulation share | A meaningful share of sampled runtime shows hands working on objects | Random sample of clips with runtime breakdown |
| Hand and object visibility | Hands and handled objects are in frame and in focus during key steps | Frame-level audit of the sample |
| Procedure mapping | Clips map to SOP, procedure or job-code IDs | Catalog export with IDs and revision dates |
| Task coverage | Your target tasks appear, performed by more than one person | Title list with tasks and durations |
| Variation | Objects, layouts, sites and performers vary | Counts by site, performer and product line |
| Working realism | At least some clips show work at normal pace in live settings | Production notes; sample comparison |
| Language quality | Narration or captions describe what is on screen when it happens | Transcripts for sampled clips |
Technical
| Check | Pass if | Evidence to request |
|---|---|---|
| Source masters | Original files exist at source resolution and frame rate | File list with codec, resolution and frame rate |
| Timing | Constant frame rate and usable timestamps | Media metadata for sampled files |
| Editing | Cuts, overlays and picture-in-picture do not hide steps | Edited and raw versions of one title |
| Camera notes | Camera type and mount are known for at least the floor footage | Production records or crew notes |
Rights and provenance
| Check | Pass if | Evidence to request |
|---|---|---|
| Ownership per title | The business owns each title or can license it; vendor titles are marked | Production contracts and course licenses |
| People on camera | Releases or notices cover employees, contractors and any actors | Release forms and notice text |
| Third-party content | Music, logos, customer sites and OEM footage are identified | Asset or clearance log |
| Confidential content | Proprietary processes and screens are identified for review | Supplier's own review notes |
For a full delivery-level test of segmented clips, use the acceptance criteria for demonstration video.
Red flags
- Offered by total hours with no title-level list.
- Most runtime is slides, talking heads or screen recordings.
- Masters are lost and only compressed LMS streams remain.
- Vendor courses are mixed in without labels.
- Nobody can say who appears on camera or what they agreed to.
- Footage was shot at customer premises or shows customer products.
- The only sample on offer is a highlight reel.
Illustrative sample scorecard
# Illustrative only — synthetic results for a fictional archive
archive: l_and_d_archive_supplier_042 (titles: 640, runtime: 210 h)
sample: 50 titles drawn at random by the buyer
results:
manipulation_share: 0.31 of sampled runtime # buyer threshold 0.25 -> pass
hands_visible_key_steps: 0.58 of key steps # threshold 0.60 -> fail (near miss)
procedure_ids: 41 of 50 titles mapped # pass
target_task_coverage: 7 of 9 target tasks present # partial
masters_available: 44 of 50 # pass
vendor_titles: 9 of 50, all labeled # pass; exclude or license separately
releases_on_file: 30 of 50 titles # fail; resolve before licensing
decision: license floor-footage subset only, after release review; re-sample hand visibility
Preparation a passing archive still needs
Even a strong archive arrives as edited lessons, not training data. Expect to select relevant titles, cut them into task and step segments, align narration to steps, add captions where narration is missing, redact faces and screens, and record per-clip metadata and rights.
Public instructional datasets show the target structure and the usual gaps. COIN annotates step-level descriptions with temporal boundaries across 180 tasks in 12 domains (Tang et al., 2019), and YouCook2 annotates procedure segments with boundaries and descriptions (Zhou et al., 2017). CrossTask lists ordered steps for each task but annotates temporal boundaries only for its primary tasks (Zhukov et al., 2019), which mirrors the partial labeling typical of business archives.
Narration alone is a weak label. HowTo100M used automatically transcribed narration with no manual annotation (Miech et al., 2019), and later work found much of it misaligned with what is on screen. If you need step-aligned language, budget for alignment; instruction-paired demonstration datasets covers how to source it.
Questions to ask the archive holder
- Can you export a title list with duration, task, site, production date and procedure ID?
- Which titles were produced in-house, and which by vendors or equipment makers?
- Do you keep original camera files, or only the edited versions in the LMS?
- Who appears on camera, and what releases or notices did they sign?
- Were any titles filmed at customer sites or with customer products?
- Is narration scripted voiceover or recorded live during the task?
- Are there unused takes, B-roll or raw floor footage behind the edited lessons?
- How do titles relate to current procedures; which ones are out of date?
Permissions and provenance
Training video is often treated as internal content, so its releases may cover training employees but not licensing to an outside company for AI use. Ask whether the notice given to people on camera reaches that far, and how former employees are handled.
Vendor courses often remain the vendor's property even when the business paid for them, so check each license. See worker consent and identifiable people for the diligence questions. This is not legal advice.
Structured course material without video, such as quizzes and assessments, is a different asset; see LMS training courses. Recordings made to certify trainees, which carry proficiency judgments, are covered in vocational training and certification recordings.
Discuss your requirements
If a training archive looks promising, describe the target tasks and the checks it must pass to SourceX. SourceX looks for US businesses that hold suitable material, checks the data and the supplier's licensing permissions, and coordinates pricing, allowed uses and delivery in a license. It follows up about potential sources after reviewing the request; a request does not guarantee a match. If the archive fails on viewpoint or coverage, compare existing footage vs commissioned recording.
More context: the workplace task demonstration hub, the humanoid training data guide, and the catalog page on licensing training materials and LMS content.
Explore sourcing and licensing options with SourceX
Sources
- COIN: A Large-scale Dataset for Comprehensive Instructional Video Analysis — step-level annotation structure for instructional video.
- Cross-task weakly supervised learning from instructional videos — ordered step lists with partial boundary annotation.
- Towards Automatic Learning of Procedures from Web Instructional Videos — procedure segments with boundaries and descriptions.
- HowTo100M: Learning a Text-Video Embedding by Watching Hundred Million Narrated Video Clips — narration as an unaligned, unannotated label source.