Short answer
Web-scale instructional video offers breadth: HowTo100M holds 1.22 million narrated videos across 23,000+ tasks. But its narration is automatic speech transcription, often misaligned with what is on screen, and rights in the underlying videos are a separate question. Licensed business footage is narrower and needs relabeling, but has a defined grant, a known site and consent you can diligence.
The choice in practice
A team building video-language or task-understanding models can start from public instructional video or from footage licensed from operating businesses. The two differ on four axes: how well text matches pictures, who holds the rights, how close the setting is to real work, and how much of your target task set appears at all. Neither provides robot actions. For what either can feed in a robot stack, see what human demonstration video can support.
What web instructional video offers
HowTo100M contains 136 million clips from 1.22 million narrated instructional videos covering more than 23,000 tasks. Its captions are automatically transcribed narrations with no extra manual annotation (Miech et al., 2019).
That design explains its weakness. Later work reports that narration is often misaligned with the visuals: one estimate was that about 50% of clip-narration pairs are not aligned, and another found about 30% visually aligned. The Temporal Alignment Networks paper builds a model that decides whether a narration sentence is alignable at all and where, and releases HTM-Align, a manually curated 10-hour, 80-video subset with sparse temporal annotations (Han et al., 2022). People narrate what they plan to do, or talk about something else, so words and actions drift apart.
Viewpoint is a second issue. Instructional videos are edited, usually third-person, and often show demonstrations staged for an audience rather than work at a production pace.
Rights: a separate question from availability
Being able to download a video does not mean you hold a grant to train on it. Questions include who owns each video, what the platform terms say, what the uploader agreed to, and who appears in it. The U.S. Copyright Office's pre-publication report examines where copying may implicate rights in AI development, analyzes fair use, and discusses licensing options (U.S. Copyright Office, 2025). It does not tell you whether a given video collection is safe to use. Dataset labels also mislead: an audit of text datasets found license categories often more permissive than the author's (Longpre et al., 2023). Check public dataset commercial license checks. This is not legal advice.
What licensed business footage offers
Licensed footage comes from a business that can state what was recorded, where and under what consent, and can grant defined uses in a written license. The footage is often closer to real work: real tools, clutter, pace. Its limits are different: narrower task coverage, viewpoint fixed for another purpose, and little or no narration. See existing footage vs commissioned recording and corporate training video archives.
Existing footage usually needs new labels; see annotating existing footage. Licensed footage still has permission questions of its own: workers, customers, third-party content and confidential information. See worker consent and identifiable people.
Illustrative comparison
| Dimension | Public instructional video | Licensed business footage |
|---|---|---|
| Scale and task breadth | Very large, 23,000+ tasks in one dataset | Limited to what the business does |
| Text-video alignment | Automatic transcripts; alignment reported as weak | None unless annotated; can be commissioned |
| Viewpoint | Edited, mostly third-person | Fixed or body-worn, depends on source |
| Setting realism | Staged, audience-oriented | Real work, production pace |
| Rights | Per-video, often unclear | Single licensor, written grant |
| Consent | Uncertain for people shown | Can be diligenced |
| Labeling effort | Alignment filtering needed | Segmentation and labels needed |
Using them together
A practical stack uses public video for broad pretraining where rights allow, and licensed footage for domain-specific fine-tuning and evaluation sets. Whichever you use, a robot still needs robot data for control; see when video is not enough.
Questions to ask
- For public video: who owns each video, and what grant do we hold, in writing?
- How were captions produced, and has alignment been measured on a sample?
- For licensed footage: what consent covers the people shown, and for which uses?
- What viewpoint, resolution and frame rate were recorded?
- Can the licensor grant training use, and are derivatives and model weights addressed?
- What labels exist, and who made them?
- Which tasks and sites are covered, and which are missing?
- Can we hold out a test set by site or operator?
Practical screening steps for each source
For public video, draw a sample of a few hundred clips and have a person judge two things: does the narration describe what is on screen at that moment, and is the viewpoint usable for your task? That gives your own alignment rate, rather than an estimate from another paper's subset. Then trace rights for a sample of the sources, not just the compilation.
For licensed footage, review sample clips for viewpoint, resolution, occlusion by gloves or tools, and how much of the task is on screen. Ask for the consent basis in writing and check how redaction affects the images you need; blurring faces does not by itself resolve permission questions. See redaction in demonstration video.
Where narration helps
Narration is not a defect when it is aligned and purposeful. Narrated demonstrations recorded for training, where a worker or annotator speaks steps as they happen, give text tied to timing; see narrated demonstrations. The contrast with transcribed web narration is that the former is designed for alignment.
Limitations of both sources
Neither web instructional video nor licensed business footage contains robot actions, joint states or contact forces, so neither can train a humanoid's control policy on its own; both need robot data, teleoperation or instrumented capture alongside them. Web video was usually filmed to teach people, so it skips steps viewers already know, cuts away during the hardest moments and frames shots for clarity rather than a robot's viewpoint. Licensed footage was recorded for training, compliance or monitoring, so its viewpoints, coverage and labels were not designed for model training either; expect gaps in hand visibility, missing step boundaries and few failure examples.
Narration and captions are weak labels in both cases: spoken steps often do not line up with what is on screen at that moment. Budget for annotation and alignment, and check a sample before committing to volume. Rights clearance for licensed footage reduces legal uncertainty, but it does not make the content more representative of your target tasks; coverage still has to be checked against your task list.
Discuss your requirements
If licensed footage fits your plan, describe the tasks, viewpoint and rights you need in a request to SourceX. SourceX investigates whether US businesses hold suitable material, checks the supplier's licensing permissions and follows up about potential sources; a request does not guarantee a match. See the demonstrations hub and the humanoid data guide.
Discuss your humanoid training-data requirements with SourceX.
Sources
- HowTo100M — scale and automatic narration.
- Temporal Alignment Networks for Long-term Video — narration alignment and HTM-Align.
- The Data Provenance Initiative — license-label mismatch.
- Copyright and Artificial Intelligence, Part 3 — copying, fair use and licensing considerations.