Short answer
Inspector video records how skilled people pick up, rotate, tilt and light a part to expose defects, and where their eyes go while they do it. It can support inspection-procedure understanding, view-planning research and proficiency evaluation. It does not contain robot actions or contact forces, and it is not a labeled defect archive, so it complements image data rather than replacing it.
Why handling behavior is its own data
A defect image tells you what a flaw looks like once it has been found. It says nothing about how the inspector found it. Skilled inspectors do not stare at a part from one angle. They rotate it under a raking light, run a thumb along an edge, tilt a glossy surface to catch a reflection, compare against a reference sample and sometimes reject on a feel that never shows up in a photo.
That sequence is what a mobile or humanoid inspection system has to reproduce or plan around: which views, in which order, with which lighting change, and when to stop. Researchers working on active inspection by humanoids need examples of the search behavior, not only the outcome. The cluster overview at Inspection, Defect and Failure Data places this page next to the image-centered pages, and the broader map is in the humanoid training data guide.
What this data can support, and what it cannot
Inspector video can support:
- Procedure and step understanding. Segmenting a check into pick-up, orient, illuminate, scan, compare, decide, and set down.
- View planning research. Learning which orientations and lighting changes inspectors use for a given part family and defect class.
- Gaze and attention models, if eye tracking was recorded.
- Proficiency and technique comparison, if several inspectors perform the same check. Ego-Exo4D includes a proficiency-estimation benchmark over skilled activity, which shows the framing is established for other domains (Grauman et al., 2023).
- Video-language pretraining, when narration or expert commentary is attached.
It cannot support:
- Robot control. Human video has no joint states, commanded actions or torques. Turning it into manipulation data needs retargeting, inferred actions or co-training with robot data. See the human-to-humanoid embodiment gap.
- Contact and grip force, unless gloves or tools were instrumented.
- Defect detection training, on its own. Frames in a handheld video are rarely resolved, calibrated or labeled well enough, and defect classes are usually not annotated.
- Proof that the shown technique matches the written standard. Inspectors often deviate from the work instruction, and the video will show what happened, not what was approved.
Existing footage or new recording
Existing footage of inspection stations is common. Quality teams record training clips, audit sessions and incident reviews, and some stations have fixed cameras. These recordings were made for another purpose, so viewpoint, resolution and labels were not designed for training. The trade-offs are set out in existing footage versus commissioned recording.
Commissioned capture lets a buyer fix the camera layout, require a fixed protocol and design consent in from the start, but it takes time, site access and a briefed set of inspectors. Most buyers end up with a mix: existing clips to learn what variation exists, and a smaller purpose-built set for the part families and defects that matter. Describing what you need at SourceX's buyer request form is the first step; SourceX's role is sourcing, rights review, licensing and delivery coordination, and specialist capture partners can be engaged case by case.
Illustrative capture plan
The values below are synthetic and are set by the buyer.
| Element | Illustrative setting | Why it matters |
|---|---|---|
| Egocentric camera | Head-worn, wide field of view | Shows what the inspector sees and reaches for |
| Exocentric cameras | Two fixed, covering bench and hands | Resolves hand-part contact and occlusion by the head |
| Close-up camera | Fixed over the part, macro-capable | Records the surface under the inspection light |
| Gaze | Eye tracking on, calibrated per session | Supports attention analysis |
| Lighting log | Light type and angle per check | Explains why a defect became visible |
| Part ID | Barcode or serial in the first frame | Links video to disposition records |
| Reference | Golden sample visible when used | Shows comparison behavior |
| Narration | Optional think-aloud, separate track | Gives language for why a part was rejected |
Project Aria is one published example of a head-worn research device that combines an RGB camera, SLAM cameras, eye-tracking cameras and IMUs (Engel et al., 2023). A buyer does not need that specific device, but the sensor mix is a useful reference for what a head-worn rig can record. Clock synchronization across the egocentric, fixed and gaze streams must be verified, not assumed.
Labels worth asking for
Raw video becomes useful when it is linked to what the inspector decided. Ask for:
- Part and lot identifiers, linked to the recorded disposition (accept, rework, scrap), which is a different label from defect type; see defect severity and disposition labels.
- Time-stamped steps: pick-up, orientation changes, lighting changes, comparison to reference, decision.
- Defect location in the frame at the moment of detection, where the inspector points or announces it.
- Inspector experience band, as a coarse category rather than a name.
- Which work instruction or acceptance standard applied.
Disposition labels from different inspectors on the same part will not always agree, which affects any model trained against them. That is covered in inter-inspector agreement and label noise.
Limitations and additional preparation
Gloves, magnifiers and body position hide the interaction that matters most. Head-worn video blurs during fast part rotation. Fixed cameras lose the surface detail the inspector sees. A single station with a few inspectors will overfit any model to those people and that lighting.
Preparation usually includes calibrating cameras, trimming idle time, aligning streams, redacting or masking faces and badges where the recording design allows, and cross-checking labels against records. Redaction can obscure hands or parts, which reduces training value, and it does not settle whether a recording may be used. Faces, voices, badge text and visible product designs each raise permission questions; see worker consent and identifiable people in workplace recordings.
Evaluating a sample before you commit
Ask for a sample session rather than a showreel. Open one clip and check the basics: Can you see the inspector's hands and the part surface in the same frame? Is the lighting change visible? Does the close-up camera resolve the feature that the inspector is looking at? Is there a timestamp or marker that lets you align the egocentric and fixed views?
Then check coverage. A set of two hundred clips from one inspector on one part family teaches less about technique than a smaller set across several inspectors, shifts and part geometries. For parts that need turning in the hand, multi-view inspection imagery shows how fixed multi-angle capture complements handheld inspection, and the question of which defects the inspection was looking for is answered by the image-centered pages such as sourcing defect image datasets.
Finally, check the baseline of variation. Skilled inspectors do the same check differently, and some of that difference is real expertise while some is habit. If the buyer plans to compare techniques or estimate proficiency, the recordings should include inspectors of different experience performing the same check on the same parts, with the experience band recorded.
Questions to ask suppliers
- Which inspection task do the clips show, and for which part families and defect classes?
- How many distinct inspectors, stations and sites are represented, and are the same inspectors repeated across clips?
- What camera positions were used, and can you share a sample session with all streams aligned?
- Is each clip linked to a part ID and a recorded disposition, or only to a date?
- Did inspectors know they were recorded, and what consent covers training use?
- Were any clips edited, for example to remove rejected parts or pauses?
- Which clips contain customer part markings, drawings or confidential screens?
- Can the supplier grant rights for the video as well as for the part imagery inside it?
Discuss your requirements
If you are weighing existing inspection footage against a purpose-built recording, describe the inspection task, viewpoints and permitted uses to SourceX. SourceX reviews the requirement and follows up about potential sources, checks the data and the supplier's licensing permissions, and coordinates the license and delivery. A request does not guarantee a match. The catalog page on first-person video of skilled manual work describes a related physical-world path.
Describe the workplace demonstration data your robotics project needs
Sources
- Ego-Exo4D: Understanding Skilled Human Activity from First- and Third-Person Perspectives — paired egocentric and exocentric capture of skilled activity with a proficiency-estimation benchmark.
- Project Aria: A New Tool for Egocentric Multi-Modal AI Research — example head-worn sensor set (RGB, SLAM, eye tracking, IMUs).