Short answer
Buying physical defect samples means acquiring real parts with known flaws, plus good reference parts, so you can image them on your own rig under controlled conditions. It gives repeatable evaluation and own-sensor data. It does not give ready-made labels, and it cannot show handling, so robot manipulation still needs robot or teleoperation data.
Why buy parts instead of images
Licensed images arrive already shaped by someone else's camera, lighting and compression. If your humanoid will inspect with a particular sensor at a particular standoff, those images are a proxy at best. A physical sample lets you photograph, scan or probe the same flaw many times, with your sensor, in the conditions you will deploy, and with the flaw presented in the poses a robot would encounter.
Physical samples also let you repeat evaluations. If a model changes, you can image the identical parts again and compare results without worrying that the underlying data shifted. That repeatability is hard to get from archives, where each image is a single moment.
The trade-off is effort. You need storage, handling procedures, an imaging rig, labeling and records. This page covers sourcing the parts. Archive-based routes appear in defect image datasets. The hub is inspection, defect and failure data and the wider map is in the humanoid training data guide.
What to buy
- Defective parts with known defect types, locations and, ideally, a disposition from the maker's own inspection.
- Matched good parts from the same lot, so the model sees the nominal appearance.
- Borderline parts near the accept-reject line, which are the hardest cases; see borderline samples.
- Reference or golden samples, covered in golden samples and reference images.
- Documentation, such as drawings, specifications or the inspection record for each part.
Typical suppliers are manufacturers holding scrap, rework or quarantine stock, and distributors who handle returns. Scrap is sometimes the cheapest source of defects, but scrap bins mix causes and lose traceability unless someone tags each piece.
Why imaging conditions matter at your end
Public benchmarks show that conditions drive difficulty. The MVTec AD 2 dataset contains eight scenarios and more than 8,000 high-resolution images, and its harder cases include small defects in large images, defects at the image border, transparent or reflective objects and overlapping bulk goods; one test split was captured under varied lighting conditions (Heckler-Kram et al., 2025). The takeaway for sample buyers is about method, not content: if you re-image parts, you control lighting, standoff and background, and you can deliberately vary them to test robustness. Be aware that the dataset's license terms may limit commercial use, so it is a reference for design, not necessarily a training source.
What re-imaged samples can support
- Evaluation sets under your own sensor and lighting, with ground truth known from the parts.
- Controlled variation studies, such as the same defect at several angles, distances or illumination levels, logged against imaging conditions metadata so later analysis can separate defect effects from lighting effects.
- Own-sensor training data for classes where you hold enough physical pieces. Hold out whole parts, not images, when building defect detection evaluation sets.
- Calibration checks between lab imaging and field imaging.
Limits: what this approach cannot support
Physical samples are not labels. A part arrives with at most a tag or a record. Region-level annotation, severity grading and taxonomy mapping are work you must do or commission. See defect annotation formats.
They cannot show how a defect arises in production. A sample records the final state, not the process variation that produced it. Defect frequency in your purchased set is set by what you chose to buy, not what a line produces.
They cannot teach handling, and parts imaged in a fixture do not show how an object behaves in a gripper; see object recognition versus manipulation data. Holding, turning and presenting a part to a camera are robot skills that need robot or teleoperation data. Samples are useful props for collecting such data if you build it yourself, but they are not that data.
Counts are limited by what exists. Rare defects may have few physical examples, which is the reason to read rare defects and class imbalance and synthetic versus real defects before deciding whether augmentation fills gaps.
Illustrative sample catalog
The following is an illustrative catalog layout with synthetic values, not an existing inventory.
| Field | Example |
|---|---|
| sample_id | smp_0034 |
| part_family, lot | housing_C, lot_2209 |
| status | defective / good / borderline / reference |
| defect_type, location | porosity, boss_left, "as tagged by supplier" |
| source_record | supplier inspection note ref_77 |
| condition_notes | surface oxidized after storage |
| storage | bagged, desiccant, shelf_B3 |
| imaging_runs | run_001 (rig_A, ring light), run_002 (rig_A, dome light) |
Keep a record of every imaging run. Storage can alter parts: oxidation, handling marks and adhesive residue may appear after purchase and confuse later labels.
Quality and acceptance criteria
You set the thresholds. Checks on receipt: each part can be matched to its tag; claimed defects are visible or measurable on inspection by a reviewer; good and defective parts come from the same lot or process; packaging prevented new damage; the quantity per defect type meets the plan you have made for your own evaluation design. Any sizing arithmetic should use arbitrary inputs of your own, for instance counting parts times imaging poses times lighting setups, and not a recommended minimum.
Permissions and provenance
Physical goods raise different questions from licensed data. Buying a part transfers ownership of the object, but it may not transfer rights in the design, marks or software on it, and it may not let you publish images. Parts may carry serial numbers, customer markings or export-controlled content. Ask what the sale terms allow regarding imaging and use of the resulting images for training, and whether customer-owned parts are included. Shipping may involve restrictions. This is not legal advice.
Questions to ask suppliers
- Are defective parts tagged with defect type and location, and by whom?
- Can matched good parts from the same lot be included?
- How were parts stored, and for how long, before sale?
- What does the sale permit regarding imaging and using images for model training?
- Do parts include customer designs, markings or serial numbers?
- How many examples per defect type are available, and how were they selected?
- Are drawings or inspection records available for each part?
- Are there export, safety or hazardous-material constraints on shipping?
- Could the supplier provide a small trial lot first?
Discuss your requirements
If you want to discuss sourcing physical samples, describe the part family, defect types and intended use 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, and SourceX follows up about potential sources. The catalog page on images and inspection photos covers the archive route.
Explore sourcing and licensing options with SourceX
Sources
- The MVTec AD 2 Dataset: Advanced Scenarios for Unsupervised Anomaly Detection — harder imaging scenarios and varied-lighting test split, as a design reference.