Short answer
Active inspection means a robot changes what it sees by moving the part or itself: picking it up, rotating it, tilting it under light, opening a panel. Fixed-camera defect images show a single controlled view, so they support detection but not choosing viewpoints or handling parts safely. Humanoid inspection needs multi-view imagery plus handling and viewpoint data, usually from new collection.
Inspection is manipulation
Most inspection datasets assume the part arrives in a known pose in front of a fixed camera. The task is passive: classify or segment what appears. A humanoid working in a workplace rarely gets that. A part sits in a bin, a panel is closed, a label faces away, a weld is behind a bracket. To see the defect, the robot has to act.
That turns inspection into a loop: look, decide where the evidence is missing, move, look again. The model needs three kinds of competence that passive datasets do not teach: choosing the next view, executing the handling motion without damaging the part, and combining observations from several views into one judgment.
What fixed-camera defect data can and cannot support
| Capability | Fixed-camera defect images | What else is needed |
|---|---|---|
| Recognizing a defect in a good view | Yes | Labels and acceptance definitions |
| Knowing which view reveals a defect | No | Multi-view data tied to defect location |
| Picking up and presenting the part | No | Manipulation demonstrations on the robot |
| Avoiding handling damage | No | Contact and force information; see handling-induced damage |
| Using lighting angle or tilt to reveal flaws | Partly | Captures under varied lighting with recorded geometry |
| Judging a part in place, in clutter | No | In-situ images from the robot's own cameras |
Passive datasets stay useful. They define what a defect looks like and provide the classification and segmentation labels. They do not show the process that gets the sensor to the right place. For choosing between normal-only and labeled-defect training data, see anomaly detection vs supervised defect data.
What multi-view and human-activity datasets show
Real-IAD offers 150K high-resolution images of 30 objects captured from multiple views for industrial anomaly detection (Wang et al., 2024). Multi-view capture exists because a defect visible from one angle may be invisible from another. It is still a fixed capture protocol: the views are predetermined, not chosen by an agent, so it supports multi-view fusion but not viewpoint planning.
Ego-Exo4D recorded skilled activities, including bike repair, with a head-worn camera and multiple exterior cameras at the same time, and defines benchmarks for fine-grained activity understanding and proficiency estimation (Grauman et al., 2023). It is a reference for the data shape that active inspection by humans would produce: what the person sees, what the room sees, and how skilled they are. It is not an inspection dataset and holds no robot actions.
What active inspection data looks like
Some inspection questions are about object state (open or closed, seated or not), which is its own labeling topic; see object state recognition.
A buyer designing a collection should think of each inspection as an episode with these elements:
- Initial state: where the part is, how it is oriented, what occludes it.
- Handling actions: grasp, lift, rotate, tilt, open, reposition. For humans, hands and part pose; for robots, joint states and commanded actions.
- Views: every image, with camera pose, lighting setting and the part's pose at capture time.
- Findings: defect labels linked to the view that revealed them.
- Decision: accept, rework, reject or escalate.
Human inspectors can generate this from head-mounted video. It shows when they pick up, turn and re-look, which is a rich source of viewpoint-selection behavior, though not of robot-executable actions. Existing manual-inspection recordings are covered in manual inspection procedure video. Robot-executed inspection needs robot episodes.
Where the data comes from
Existing assets. Quality departments may hold multi-view image sets from automated stations and photographs of defective parts. These support the perception side. Fixed workplace cameras rarely show the inspector's hands closely enough to show handling; see fixed-camera footage limits.
Commissioned human recording. Inspectors wear head-mounted cameras and work in front of fixed cameras while a lighting and tilt protocol is followed. This produces both viewpoint choices and handling in the same recording.
Robot or teleoperated capture. Needed when the model must execute the handling. A specialist provider with the target embodiment is the usual route.
Synthetic views. Rendered views from 3D models can augment multi-view training but do not capture handling or realistic lighting artifacts; see 3D surface inspection data.
Illustrative capture protocol for human active inspection
# Illustrative only - synthetic values
task: inspect_housing_for_surface_and_edge_defects
subjects: inspectors, operator_A12 to operator_A19
cameras:
head_mounted: 1080p, 30 fps, intrinsics provided
fixed: 2 views, calibrated, time-synced
lighting: ring light, raking light (angle logged), ambient
required_views: [top, bottom, 4 sides, edge_raking]
part_pose: logged via fiducial markers
labels:
per_view_findings: defect_class + mask
handling_events: pick, rotate, tilt, put_down (start/end times)
decision: accept | rework | reject | hold
viewpoint_reason: free text or coded
exclusions: no force data; no robot actions
The protocol yields human viewpoint behavior and multi-view evidence. Converting it to a humanoid's inspection routine still requires retargeting or robot data, and the usual embodiment gap applies.
Specifying the purchase
- State which part of the loop the data must teach: view selection, handling, multi-view fusion or the decision.
- Specify camera pose and lighting metadata for every image; see multi-view inspection imagery.
- Require defect labels linked to views, not only to parts.
- Decide whether failures and handling damage should be included.
- Keep an evaluation set of parts and defects not used in training.
Questions to ask suppliers
- Are images captured at fixed stations, by inspectors, or by robots?
- Is the camera pose recorded for each image, and is it calibrated?
- Do multiple views of the same part share an identifier?
- Are lighting settings logged and varied?
- Is the handling visible, or only the part in the final presented pose?
- Can defect labels be traced to the view where the defect is visible?
- Do inspectors wear gloves or use tools that hide the part?
- Are rejected parts available for re-imaging?
Discuss your requirements
If you are specifying inspection data for a robot that must handle and re-view parts, describe the parts, views and handling to SourceX. SourceX investigates whether US businesses hold suitable material and, if no existing asset fits, can scope a commissioned collection of new recordings; a request does not guarantee a match. See the humanoid training data guide, the inspection and defect data hub, and the catalog entries for images and inspection photos and egocentric video of manual work.
Discuss your humanoid training-data requirements with SourceX
Sources
- Ego-Exo4D: Understanding Skilled Human Activity from First- and Third-Person Perspectives — synchronized first- and third-person recordings of skilled activity.
- Real-IAD: A Real-World Multi-View Dataset for Benchmarking Versatile Industrial Anomaly Detection — multi-view industrial anomaly images.