Short answer
Parts identification data pairs images or video of components as they appear at a workstation, in bins, trays, kits or fixtures, with part numbers, pose or location labels and presentation context. It supports detection, identification and pose estimation. It does not show how a part behaves when grasped or inserted, so manipulation still needs robot or demonstration data.
The buyer's problem
An assembly robot or humanoid has to know which part it is looking at before it can reach for it. At a real station that is harder than recognizing everyday objects. Parts are small, often metallic and low in texture, many look alike across revisions, and they arrive in trays, bins, blister packs or loose heaps with varying orientation. A recognizer trained on clean catalog renders or isolated product photos can fail on a glare-covered tray.
This page covers sourcing images and video of parts as presented at stations, and the metadata that determines whether they are useful. For the distinction that governs what such data can do, see object recognition data versus manipulation training data.
What published benchmarks show about the difficulty
Two public resources illustrate what a well-built pose dataset includes and why industrial parts are hard.
T-LESS contains 30 industry-relevant texture-less objects, many of which are symmetric or similar to each other, and some of which are parts of others. Each scene was captured with three synchronized sensors (structured-light RGB-D, time-of-flight RGB-D and high-resolution RGB), with roughly 39,000 training and 10,000 test images per sensor, and each object has a CAD model and a reconstructed 3D model (Hodan et al., 2017).
BOP defines a benchmark for 6D pose of rigid objects from a single RGB-D image. It brings eight datasets into a unified format, adds a pose-error function that handles ambiguous poses, and evaluates 15 methods (Hodan et al., 2018).
Three lessons carry over to buying. First, symmetry and similarity mean that labels must say how ambiguity is handled. Second, a 3D model per part is part of the dataset, not an extra. Third, multiple sensors and a shared format exist because capture choices change results. Those are published benchmarks, not a size recommendation for your data; requirements depend on your architecture, parts and use.
Where station-level parts data exists
- Manufacturers' vision and inspection systems store images of parts at kitting, feeder and inspection stations, usually with a result flag.
- Process-monitoring cameras capture the workstation, but rarely close enough to identify small components. See process monitoring camera footage.
- Bin-picking cell logs from existing robot installations hold depth and RGB images with pick outcomes. See industrial bin picking of parts and bin picking image datasets.
- Product and part master data from PLM or ERP systems provides part numbers, revisions and CAD, which can serve as the label vocabulary. See PLM and ERP part master data.
- Commissioned capture at a consenting station with chosen cameras, lighting and labels, for buyers who need pose and presentation variation that archives lack.
Presentation context to ask for
The same part looks different depending on how it is presented. The table lists context a good package records.
| Presentation | What varies | Label or metadata needed |
|---|---|---|
| Loose bin | Heap depth, occlusion, orientation | Part ID, visibility, pose or mask |
| Sorted tray or blister | Fixed pockets, lid reflections | Pocket ID, part ID, empty/filled |
| Kit tray | Mixed parts, sequence | Kit ID, expected vs actual contents |
| Fixture or jig | Constrained orientation | Fixture ID, seated or unseated |
| Feeder or conveyor | Motion, blur | Frame timing, part ID |
| Operator hand | Part held, partially hidden | Hand visibility, grasp type |
Lighting, camera height, sensor type and surface finish also matter. Add them as metadata fields so you can stratify evaluation by condition.
What the data can support, and what it cannot
Station-level parts images can support:
- Detection and identification of parts and variants, including checks that the right part is in the right pocket.
- Pose estimation when pose labels or CAD alignment are provided.
- Grasp planning inputs (candidate pick points from pose and geometry).
- Anomaly and mix-up detection where wrong or missing parts are labeled.
They cannot support:
- Handling behavior. Images show nothing of how a thin clip flexes, how a connector latches, or how far a part slips when grasped. That needs contact data from robot or instrumented capture.
- Robot action spaces. No joint states or gripper commands.
- Generalization to unseen revisions unless the dataset includes them.
If the same footage shows an operator picking from the bin, it adds a human demonstration, not a robot one. See what human demonstration video can and cannot do.
Illustrative package for parts identification
# Illustrative only - synthetic values, not an available dataset
package: station_parts_id_v0
station_types: [kit_tray, loose_bin, fixture]
sensors: [rgb_global_shutter, tof_depth] # buyer-specified
calibration: intrinsics + extrinsics per camera, hand-eye matrix if robot present
annotations:
part_id: matches plm_part_number
revision: letter or number
pose_6d: where CAD alignment exists
mask: instance segmentation
visibility: fraction visible
ambiguity_flag: symmetric or near-identical pair
models: CAD per part, units in mm
splits: by station and by day; held-out stations for evaluation
rights: image and CAD license terms stated separately
Quality and acceptance checks
- Sample images at each station type and confirm label correctness against physical parts or master data.
- Check that symmetric or near-identical parts are labeled consistently and the handling is documented.
- Confirm that the held-out split is by station, shift or lot, not random frames from the same sequence.
- Verify that CAD models match the revision shown in the images.
- Count images per part and per condition so rare parts are not an afterthought.
Limitations and additional preparation
Existing images were usually taken for inspection or logging, so they may be low resolution, poorly aligned, or lack pose. Labels often reflect pass/fail, not part identity. CAD and part numbers may be confidential, and a manufacturer may only release images, not drawings, so check what the supplier can license; see manufacturing data confidentiality. A commissioned capture can be designed with pose labels, but it still needs site access, calibration work and consent from people in frame.
Presentation, lighting and sensor choices
Station images look different from benchmark scenes. Shiny parts return specular glare under overhead lights, dark rubber parts vanish against black trays, and clear lids add reflections. The published datasets above deliberately used several sensor types and varied lighting for these reasons. When you specify a commissioned capture, state the sensor classes you expect your robot to carry, because a model trained on one depth technology may behave differently on another.
Ask also about the time dimension. Parts at a feeder or a moving conveyor are seen in motion, while kit trays are static. Video with part IDs per frame supports tracking, whereas isolated stills do not.
Questions to ask suppliers
- Which station types and presentation styles are in the images?
- How were part IDs assigned: by system lookup, operator scan or manual review?
- Is there pose, mask or bounding-box labeling, and who produced it?
- Do you have CAD or 3D scans for each part, and can they be licensed with the images?
- How are symmetric or near-identical parts labeled?
- What camera, lens, lighting and mounting were used, and are calibration files available?
- Which part revisions appear, and are revision labels reliable?
- Do any images show customer markings, drawings or people?
- Can a new capture be run at a station with a defined part list?
Discuss your requirements
If you need images of parts as they appear at workstations, describe the parts, stations 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. If no existing asset fits, SourceX can scope a commissioned collection of new recordings. For the wider map, start at the humanoid training data guide or the manufacturing and assembly hub. The site's entry on images and inspection photos describes the data type, and fixturing and part orientation covers the constraint side.
Discuss your humanoid training-data requirements with SourceX
Sources
- T-LESS: An RGB-D Dataset for 6D Pose Estimation of Texture-less Objects — texture-less industrial objects, three synchronized sensors, CAD models.
- BOP: Benchmark for 6D Object Pose Estimation — unified-format benchmark and ambiguity-aware pose error.