Short answer
Fastener identification data pairs close-range images of screws, bolts, nuts, washers and clips with the attributes that distinguish them: drive type, head style, thread, length, diameter and material. Sources are distributors, manufacturers and workshops. The hard part is similarity: parts that differ by a millimeter or a thread pitch look alike, so resolution, scale references and exact labels matter more than image count.
The buyer's problem
A humanoid that assembles, repairs or kits parts must tell an M5 screw from an M6 screw, a pan head from a button head, and a lock washer from a flat one. Many such parts are small, metallic and texture-less, so appearance cues are weak. Mistakes are cheap to make and costly to find, because a wrong fastener may only fail later.
Fasteners are an extreme case of fine-grained recognition. Before buying, decide whether you need to identify exact part numbers, classes by attribute or just to sort loose parts into bins. For the broader catalog context, see industrial distributor catalogs; for tiny-part imaging constraints, see small object resolution requirements.
What a published industrial-parts dataset teaches
T-LESS provides 30 industry-relevant texture-less objects, many of which are symmetric or mutually similar, and some of which are parts of others. Data were captured with three synchronized sensors (structured-light RGB-D, time-of-flight RGB-D and a high-resolution RGB camera), with about 39,000 training and 10,000 test images per sensor and a CAD model and reconstructed 3D model per object (Hodan et al., 2017).
Four buyer lessons follow. Symmetry makes pose labels ambiguous, so your annotation scheme needs a symmetry definition. Parts that are subsets of others make "which object is this" depend on what is visible. Several sensors in one dataset let you test sensor sensitivity. And a CAD model per object lets you generate synthetic variants and check labels, which is a pattern worth requesting for hardware. T-LESS has 30 objects; a fastener catalog has thousands, so it shows method, not coverage.
Attributes that define a fastener
| Attribute | Examples | Where it comes from | Visible in an image? |
|---|---|---|---|
| Type | Screw, bolt, nut, washer, rivet, clip | Catalog | Usually |
| Head style | Pan, hex, countersunk, button | Catalog | Yes, if head faces camera |
| Drive | Phillips, Torx, hex socket, slotted | Catalog | Only at close range |
| Thread standard and pitch | Coarse versus fine | Catalog or gauge | Weakly; needs gauge or high resolution |
| Diameter and length | By standard size | Catalog or caliper | Only with a scale reference |
| Material and finish | Steel, stainless, zinc, black oxide | Catalog | Partly |
| Grade or class | Strength marking on head | Head stamp | Only at close range |
Three of those attributes (thread pitch, diameter, length) cannot be read from a single unscaled photo. Specify a scale reference or calibrated camera, or accept that the model will rank candidates and a downstream step (gauge, fixture, second view) will confirm.
Supplier types
Distributors hold catalogs with images, part numbers and structured attributes, often in the thousands. Images are usually studio shots, one angle per item, and imagery is sometimes shared across a family of sizes, which makes it a poor label for size.
Manufacturers hold CAD, drawings and standards-based specifications. These supply geometry and exact attributes but not photographs of parts in use.
Workshops, maintenance departments and assemblers hold images of bins, trays and assemblies, which show mixing, grime and orientation but seldom have per-part labels.
Existing catalog data is the realistic first purchase; commissioned photographs of a bounded part list on trays or in hands are a reasonable second. See automotive parts catalog data and electronic components recognition for adjacent part families.
Illustrative package
# Illustrative: synthetic values only
part_record:
part_id: fst_00921
attributes: {type: screw, head: pan, drive: torx_T20, thread: M5x0.8, length_mm: 16, material: steel_zinc}
images:
- {file: fst_00921_a.png, view: top, scale_ref: ruler_mm, px_per_mm: 38, background: tray_grey}
- {file: fst_00921_b.png, view: side, scale_ref: ruler_mm, px_per_mm: 38, background: tray_grey}
- {file: fst_00921_c.png, view: bin_mixed, scale_ref: none, count_in_frame: 14}
cad: fst_00921.step
confusable_with: [fst_00922, fst_00935] # near neighbors by attribute
source: distributor_catalog_export
The confusable_with field is useful: it lets you build test pairs that probe exactly the distinctions your robot cannot get wrong.
Evaluation and acceptance
- Confusion-pair test set. Build pairs differing by one attribute and report accuracy on each attribute separately.
- Scale audit. Check pixels per millimeter across the data; reject sets with no scale control if size matters.
- Label verification. Re-measure a sample with calipers and a thread gauge against catalog values.
- Mixed-bin scenes. Include images of many parts together, not only singles; see cluttered scenes for recognition.
- Pose ambiguity handling. For pose tasks, require symmetry annotations; see 6D pose annotation.
Imaging setup guidance
Fastener images from an overhead camera need enough pixels across the smallest feature that distinguishes classes. For a drive recess or a thread, that can mean tens of pixels across the feature, which for a small screw implies a close lens or a high-resolution sensor. Wide-field bin images cannot carry that detail. Many programs use a two-stage design: a wide view to find candidate parts, then a closer view from the wrist camera or a fixture to confirm identity. Ask suppliers for both kinds of image and for the camera geometry, so you can simulate your own pipeline. Lighting also matters: low-angle light shows thread and knurling, while diffuse light hides them but reduces glare on plated heads.
What this data cannot support
Identifying a fastener does not tell a robot how to drive it. Threading, torque and cross-threading avoidance are force-sensitive manipulations that need interaction data, and recognition models do not address them. The same holds for picking tiny parts from a bin, which depends on gripper size and a kits and multi-part products workflow. Catalog images may show a plated finish that differs from what is in the bin. Requirements depend on the part list and robot, and no universal image count suffices.
Permissions and provenance
Distributor imagery often belongs to manufacturers or photographers, and part-number schemes are proprietary. Ask who can license training use for both images and the attribute table. This is not legal advice.
Questions to ask suppliers
- How many distinct part numbers, and how many images per part?
- Are images unique per size, or shared across a family?
- Is a scale reference or calibration included?
- Which attributes are verified by measurement and which copied from catalog text?
- Are CAD files or drawings included for each part?
- Do you have images of mixed bins, trays or assemblies?
- Can near-neighbor parts be identified for building confusion tests?
- Who owns the photographs and the attribute data, and can both be licensed for training?
Discuss your requirements
If you can list the fastener families, sizes and the distinctions your robot must make, describe the requirement to SourceX. SourceX reviews it, follows up about potential US sources such as distributors or manufacturers, and checks the data and licensing permissions. A request does not guarantee a match. Start from the humanoid training data guide or the product and object data hub. The site's entries on product catalogs and descriptions and CAD and engineering drawings are relevant categories.
Discuss your humanoid training-data requirements with SourceX
Sources
- T-LESS: An RGB-D Dataset for 6D Pose Estimation of Texture-less Objects — T-LESS: 30 texture-less, often symmetric or mutually similar industrial objects, three sensors, CAD and reconstructed model per object.