Short answer
Ergonomic assessment videos are recordings of workers performing tasks, made to score posture, force, repetition and lifting risk. They often come with task breakdowns, timing and risk scores. They can support task decomposition, posture and motion-element analysis, but they are made for assessment, may not be full quality, and typically involve identifiable employees and consultancy ownership.
Why these recordings exist
Employers record work to judge whether it injures people. OSHA describes musculoskeletal disorders as conditions affecting muscles, nerves, blood vessels, ligaments and tendons, and says prolonged exposure to risk factors such as forceful exertion, repetitive motion and awkward or static postures can cause them; it frames an ergonomic process as ongoing (OSHA, Ergonomics - Overview). Assessors therefore film jobs, decompose them, and score the postures and loads using a method such as the Revised NIOSH Lifting Equation, which estimates the risk of low back disorders from manual lifting, including asymmetrical lifts (NIOSH, 2021).
The result is footage organized around the body: where the hands were, how far the load was lifted, how often. For a buyer, that is an unusual combination of video plus a structured description of the task, made by someone who watched it closely. The question is what such material can add to embodied-AI training, and how to evaluate it.
What a typical assessment package holds
| Element | Typical content | Value to a buyer |
|---|---|---|
| Video | Fixed or handheld clips of a job cycle, often from two angles | Task examples with posture visible |
| Job description | Task name, tools, loads, frequency, duration | Metadata for indexing |
| Task decomposition | Elements and sub-elements with start and end times | Step labels |
| Measurements | Hand height, horizontal distance, load weight, lift frequency | Quantitative context |
| Risk scores | Output of the chosen method (lifting equation, posture scores) | Weak labels for effort and posture |
| Recommendations | Proposed redesigns, before/after clips | Variation of the same task |
| Report | PDF with photos and tables | Provenance and method |
Many consultancies and EHS teams also produce time and motion study recordings, which share the decomposition into elements. Differences: ergonomic recordings emphasize the body and the risk; time studies emphasize the duration.
What the data can support, and what it cannot
Ergonomic assessments can support:
- Task decomposition. Motion elements, timing and loads for manual jobs.
- Posture and reach context. Where hands, trunk and load were relative to the worker, usable for studying embodiment differences. See the human-to-humanoid embodiment gap.
- Weight and effort metadata. Load mass and frequency that a deployment team can use when selecting tasks suited to a humanoid's payload.
- Task selection. Risk scores point to jobs where automation is attractive.
- Before and after variants. Redesigned workstations produce paired examples.
They cannot support:
- Robot actions. Posture from video is not a joint trajectory. There is no robot state or command.
- Full-quality training video in every case. Assessments are often filmed on phones or handheld cameras, from the angle the assessor needed, sometimes compressed or cropped. Check resolution, frame rate and framing.
- Contact force. Load estimates are not force traces.
- Generalization across jobs. Assessors film jobs that look risky, which skews the selection toward heavy, repetitive tasks.
For why the transfer from human to robot motion is not direct, see what human demonstration video can and cannot do. Scores are also method-dependent: a number from one tool is not a ground-truth measure of difficulty, and two assessors may score the same clip differently.
Rights and provenance
These recordings carry layered ownership.
- Employer. The manufacturer owns the facility and often the video, but it may have commissioned a consultant under a contract that gives the consultancy rights in reports and methods.
- Consultancy. EHS firms may retain ownership of analyses and templates, and their contracts may bar reuse.
- Employees. Assessment video focuses on one person's body, often in identifiable detail. Notice may have covered safety review, not licensing for AI training. See worker consent and identifiable people.
- Health-related inference. Posture and effort scores can be read as information about a person's capacity or injury risk. Ask whether any recordings relate to injury investigations or claims, which are likely to be excluded.
Blurring faces does not by itself address any of this. This is not legal advice; take these as questions for the supplier and your counsel.
Illustrative package
# Illustrative only - synthetic values, not an available dataset
package: ergo_assessments_v0
assessments: 40 jobs across 4 plants (example scale chosen by buyer)
per_job:
video: 1-3 angles, original resolution retained, frame rate stated
decomposition: elements with start/end times
measurements: load_kg, hand_height_cm, horizontal_dist_cm, lifts_per_min
scores: method_name, version, result, assessor_id (pseudonymous)
report_pdf: redacted (names, photos of people removed)
redesign_pair: before/after clip where it exists
worker: operator_A05 (pseudonymous), consent_status recorded
exclusions: injury investigations, medical information
rights: employer + consultancy permissions documented per job
Quality and acceptance checks
- Sample clips and check whether the whole body and the object are in frame through the cycle.
- Compare decomposition timestamps to the video, and report the match rate on a sample.
- Confirm which method and version produced each score, and whether the assessor is identified.
- Check frame rate and compression: cropped or heavily compressed clips limit later pose estimation.
- Confirm which clips come with usable measurements, and which are only photos.
- For any evaluation, hold out by plant and by operator.
Limitations and additional preparation
Locating the files is half the work: assessments sit in consultant archives, EHS shared drives and report attachments, rarely in one system. Linking each analysis to the correct clip needs manual effort. Some video exists only as report screenshots, and some has been edited down to key frames. Redaction of faces and names is likely to be needed, and it can reduce the visibility of the worker's posture, which is the value of the footage. If your use needs higher-quality, multi-view video with 3D hand tracking, see existing footage versus commissioned recording.
Choosing how to use the scores
Treat scores as metadata that helps you find and group clips, not as ground truth about difficulty for a robot. A high lifting index tells you the load is risky for a person; it says little about whether a humanoid with a different payload and reach can do it. Useful groupings include heavy versus light, one-hand versus two-hand, high reach versus low reach, and repetitive versus occasional. Those groupings help you select tasks and sample clips for annotation without assuming the score maps to robot difficulty.
Task element lists often derive from standard work and job breakdown sheets, which are worth requesting alongside the video.
Questions to ask suppliers
- Who recorded the assessments, and who owns the video, the scores and the reports?
- Which scoring methods and versions were used, and are the inputs recorded?
- What cameras, resolutions and frame rates were used, and are original files retained?
- Is each clip linked to a decomposition, a measurement sheet and a score?
- What notice or consent did the workers give, and did it cover licensing for AI training?
- Do any recordings relate to injury investigations, claims or medical information?
- Do consultancy contracts allow reuse of the analyses?
- Are before-and-after redesign clips available?
- What redaction would be applied, and how does it affect body visibility?
Discuss your requirements
If you want posture-focused recordings with scores, describe the tasks, loads 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. For the wider map, start at the humanoid training data guide or the manufacturing and assembly hub. The site's entry on video recordings describes the data type, and assembly demonstration video covers the broader footage category.
Discuss your humanoid training-data requirements with SourceX
Sources
- Applications Manual for the Revised NIOSH Lifting Equation (DHHS (NIOSH) Publication No. 94-110) — Revised NIOSH Lifting Equation for manual lifting risk.
- Ergonomics - Overview — musculoskeletal disorders and ergonomic risk factors.