Short answer
Trade schools, apprenticeship programs and certification bodies record students and candidates performing practical tasks, and assessors grade the result. That pairing of video with a proficiency judgment is rare and useful for quality-weighted learning and evaluation, but it shows learners, not experts, and has no robot actions, forces or consistent viewpoints. Rights involve students, instructors and sometimes minors.
Why this supplier type is different
Most workplace footage arrives with no judgment attached: you see a worker do a task and have to guess whether it was done well. Vocational settings are built around judgment. A candidate welds a joint, wires a panel, frames a wall or plates a dish, and an assessor scores it against a rubric. Some programs record the attempt for appeals, moderation or instructor feedback.
For a humanoid-data buyer, that means video paired with outcome and quality labels that are expensive to create elsewhere. It also means a different population: learners at varying skill levels, working in teaching facilities rather than production sites, under conditions shaped by the assessment. This page covers how to evaluate such material. For what published work says about human demonstrations in general, see what human demonstration video can and cannot do. For the general question of mixing skilled and unskilled demonstrations, see expert vs novice demonstrations.
What the recordings can support
Published research gives two reasons to value proficiency labels.
First, demonstration quality changes what a model learns. In a study of offline learning from human demonstrations, researchers compared six algorithms on five simulated and three real-world tasks, using datasets of varying quality; one set of 300 demonstrations from six operators was grouped into Worse, Okay and Better subsets, and results depended on demonstration quality (Mandlekar et al., 2021). Labels that separate better from worse attempts let a buyer filter, weight or stratify.
Second, proficiency is a recognized video-understanding task. Ego-Exo4D records skilled activities, including bike repair, from first- and third-person cameras, adds expert commentary, and defines benchmarks including proficiency estimation (Grauman et al., 2023). Assessment recordings are the natural industrial analogue.
| Use | What the recordings contribute |
|---|---|
| Quality-weighted imitation or reward modeling | Attempts ranked by assessors |
| Step recognition and sequencing | Rubric items that name steps and their order |
| Mistake and error detection | Common errors flagged by instructors |
| Evaluation sets | Held-out candidates scored on a shared rubric |
| Language grounding | Instructor feedback and rubric text tied to moments |
What they cannot support
- Expert behavior. Students are learning. An assessment set may contain few expert-level attempts, and "pass" often means minimum competence. Do not treat the average candidate as a target policy.
- Robot actions and forces. No joint states or commanded actions. Craft quality often depends on force and feel, such as torque on a fastener or pressure on a tool, which video does not capture. Read when video is not enough.
- Consistent viewpoints. Booth cameras, instructor phones and fixed room cameras differ by school and sometimes by session.
- Production pace and conditions. A teaching bay differs from a live site in tooling, lighting, clutter and pace.
- Coverage of your tasks. Programs teach to a syllabus; the tasks you need may be partly covered or not at all.
- Independence. Assessor scoring varies. Ask whether double marking or moderation was used before treating scores as ground truth.
Supplier types
| Supplier | Typical holdings | Watch for |
|---|---|---|
| Trade and technical schools | Practical sessions, projects, instructor demonstrations | Student consent, minors, institutional policy |
| Apprenticeship and training providers | Workplace-based assessments, logbooks, supervisor sign-offs | Employer-held footage, mixed ownership |
| Certification and licensing bodies | Practical exam recordings, rubrics, appeal files | Exam security, candidate confidentiality, retention limits |
| Equipment and tool vendors with training arms | Product-specific procedures | Third-party brands, vendor content rights |
Where a corporate learning team runs vocational courses for its own staff, the corporate training video archive guide applies as well.
Illustrative package: assessment-linked task video
# Illustrative only - synthetic values, not an available dataset
package: practical_assessment_video_v0
source: technical college program (anonymized id: provider_C04)
task_family: electrical panel wiring (cohorts 2-4)
assets:
video: fixed bench camera, 1080p, constant 30 fps
assessor_rubric: pdf per assessment, version-tagged
labels:
rubric_scores: per criterion, 0-3
overall_result: pass | refer | fail
assessor_comments: free text, timestamped where available
second_marker: yes/no per attempt
candidate_stage: year 1 | year 2 | qualified_trade
metadata: cohort_id, candidate_id (pseudonymous), assessor_id, date, bench_id
excluded: robot actions, forces, tool telemetry
rights: training and internal evaluation; no redistribution
The candidate_stage field is the one buyers most often miss. Without it you cannot separate learners from qualified tradespeople.
Permissions and provenance
Students and candidates are the people on camera, and many recordings were made for assessment or safety rather than model training. These are diligence questions, not conclusions; this is not legal advice.
- What did candidates agree to, and did it cover use by third parties for AI training?
- Are any candidates minors or in programs with special confidentiality, and do educational-record rules apply to the files or scores?
- Do assessor comments identify individuals? Names in audio and on-screen scorecards are common.
- Does the awarding body or school own the footage, or does an employer or equipment vendor?
- Are rubrics, exam tasks or standards documents copyrighted by a third party?
Redacting faces does not by itself settle consent. See worker consent and identifiable people.
Evaluation pointers
Request a stratified sample: attempts across the score range, several cohorts and several assessors. Check that hands and workpiece are visible during the graded steps, not only at the end. Compare scores to what you see; if the best-scoring attempt looks rushed or the worst looks fine, the labels are measuring something other than motion quality. Plan annotation of existing footage for step boundaries, and add success and outcome labels where the assessment outcome is too coarse.
Questions to ask suppliers
- Which tasks are recorded, and are they the practical tasks or only instructor demonstrations?
- How are attempts scored, by how many assessors, and is double marking recorded?
- Can scores be released per criterion, or only pass/fail?
- What is the mix of learner stages in the footage?
- What consent or notice covered students and candidates for each recording?
- Do any recordings include minors, and who decides on their release?
- Are camera positions fixed by room, and are they documented?
- Who owns the rubrics, task briefs and the footage, and what can be licensed together?
- How long are recordings retained, and can deletion requests be honored after delivery?
Discuss your requirements
If you are looking for skilled-trade video with quality judgments attached, describe the trades, tasks and permitted uses to SourceX. SourceX looks for US businesses that hold suitable material, checks the data and the supplier's licensing permissions, and coordinates the license and delivery. A request does not guarantee a match; SourceX follows up about potential sources. Start from the humanoid training data guide or the workplace task demonstration hub, and see the catalog page on training materials and LMS content.
Discuss your humanoid training-data requirements with SourceX
Sources
- Ego-Exo4D: Understanding Skilled Human Activity from First- and Third-Person Perspectives — skilled-activity recordings, expert commentary and proficiency estimation.
- What Matters in Learning from Offline Human Demonstrations for Robot Manipulation — outcomes depend on demonstration quality.