Industrial Bin Picking of Parts

The buyer's problem

Consumer-item bin-picking data is plentiful in public benchmarks. Industrial parts are different: machined brackets, stampings, fasteners, castings and connectors are often metallic, reflective, textureless, symmetric and prone to tangling or nesting. A model that finds cardboard boxes well can fail on a pile of identical shiny pins.

This page helps perception and manipulation leads choose among public datasets, manufacturer-held captures and commissioned scenes. It sits within the manufacturing cluster; for the general difference between recognition data and manipulation data, see object recognition data vs manipulation training data.

What public industrial datasets show

Two verified works illustrate the range of what exists, and how far it is from a specific plant's parts.

T-LESS contains 30 industry-relevant texture-less objects, many of them symmetric or similar to each other, and some that are parts of others. It was captured with three synchronized sensors (a structured-light RGB-D sensor, a time-of-flight RGB-D sensor and a high-resolution RGB camera), with about 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). It is a benchmark for 6D pose estimation, not a record of picking.

MetaGraspNet targets bin picking directly. It provides 217,000 synthetic RGB-D images across 82 article types, with annotations for detection, amodal perception, keypoints, manipulation order and both parallel-jaw and vacuum grasp labels, plus more than 2,300 real annotated RGB-D images in five difficulty levels and an unseen-object set (Gilles et al., 2022).

Both are useful for method development and for understanding what annotation types exist. Neither covers your part. Before relying on either for a commercial program, check the dataset licenses; see public dataset commercial license checks.

Why metal parts are different

Check each of these against a candidate dataset:

• Reflections and specular highlights change the apparent appearance with viewing angle and lighting, and can saturate depth sensors.
• Missing or noisy depth appears on shiny or dark surfaces for some sensor types; the failure pattern depends on the sensor technology.
• Symmetry makes pose ambiguous. Annotations must declare how symmetric poses are handled.
• Texture-less surfaces deprive feature-based methods of cues.
• Tangling and nesting (springs, wire forms, hooks, clips) means that picking one part drags another.
• Oil, coolant and packaging alter surfaces and stack behavior.
• Bin walls and dunnage limit approach directions and cause collisions.

For parts that arrive racked or kitted instead of loose, see parts presentation and identification and line-side kitting and replenishment.

Sources of supply

Existing manufacturer images are usually inspection or process-monitoring captures, framed for a different task. Check whether the bin is fully visible, whether depth exists at all, and whether any pick outcome is logged. Synthetic rendering from CAD can add volume, but the realism gap on reflective parts is a known risk; see real vs synthetic and simulation data.

What the data can and cannot support

Images, depth and pose or grasp annotations can support detection, segmentation, pose estimation and grasp candidate ranking. They do not show how a part moves when pulled, how a pile collapses, or whether a suction cup holds an oily surface. Those outcomes need physical attempts. Grasp labels computed from geometry, as in synthetic data, describe predicted success, not measured success on hardware.

A human operator reaching into a bin and a robot gripper entering the same bin differ in approach geometry, finger size and sensing. Human footage of bin picking can show which part people choose first and how they untangle parts, but it provides no robot grasp pose or action. See what human demonstration video can and cannot do for robot learning.

Illustrative capture specification

# Illustrative only - synthetic values, not an available dataset
package: bin_picking_parts_v0
parts: 12 part types (synthetic IDs part_C01 ... part_C12), CAD supplied by owner
scenes: 4 fill levels x 3 bin types x 2 lighting setups per part type
sensors:
  - structured_light_rgbd, 1280x1024 (illustrative)
  - rgb_camera, wrist or overhead mount
  calibration: intrinsics + hand-eye or overhead extrinsics per session
annotations:
  instance_masks, visible_and_amodal
  pose_6d: with declared symmetry group per part type
  grasp_labels: parallel_jaw / vacuum candidates, geometric or measured (stated)
  difficulty_level: 1-5, defined in a written rubric
splits: train / validation / test, held out by scene and by part lot
excluded: robot actions unless pick attempts are recorded on a robot

Quality and acceptance pointers

Ask for sample frames from each sensor at each difficulty level before pricing. Verify depth completeness on the shiny parts, not on the easy ones. Check annotation agreement, for example by re-annotating a random subset. Confirm that test splits are held out by scene and by part lot rather than by random image, otherwise near-duplicate frames leak across splits; see training vs evaluation datasets.

Limitations and additional preparation

• Scene coverage reflects the supplier's bins and lighting, not yours.
• CAD and part geometry are often customer-confidential; see manufacturing data confidentiality.
• Pose annotation of symmetric parts is easy to get subtly wrong.
• Sensor choice matters: a model trained on one depth technology may not transfer to another.
• Grasp labels need a defined gripper; labels for a parallel-jaw gripper do not apply to a humanoid hand.

Questions to ask suppliers

1. Which part types, finishes and materials are covered, and are they loose, nested or tangled?
2. Which depth and RGB sensors were used, and how were they calibrated?
3. How complete is depth on reflective parts, and how is missing depth marked?
4. How is symmetry handled in pose labels?
5. Are grasp labels geometric or from physical attempts, and for which gripper?
6. How were difficulty levels defined and applied?
7. Is the test split held out by scene and by part lot?
8. Who owns the CAD and part geometry, and can it be licensed?
9. Can new scenes with the buyer's parts be captured, and under what site access conditions?

Discuss your requirements

If you need industrial bin-picking imagery for specific parts, describe the parts, sensors, annotations 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. See the humanoid training data guide and the manufacturing hub, or the catalog entry for images and inspection photos.

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 industrial objects, three sensors, about 39K train and 10K test images per sensor, CAD models.
• MetaGraspNet: A Large-Scale Benchmark Dataset for Scene-Aware Ambidextrous Bin Picking via Physics-based Metaverse Synthesis — MetaGraspNet: 217k synthetic and 2.3k+ real RGB-D bin-picking images with grasp labels.