Short answer
Tool-condition recognition needs labeled images of the same tool types in serviceable, worn and damaged states, taken under controlled and in-service conditions. Workshops, tool rental operations and repair shops are plausible holders. Public manipulable-object datasets cover tool identity and pose, not wear, so condition labels, grading rules and rejection decisions must come from the supplier's own inspection practice.
The buyer's problem
A maintenance humanoid that picks a screwdriver, a torque wrench or a cutting blade has implicitly decided the tool is fit for use. People make that decision in a second: a rounded tip, a cracked handle, a chipped edge, a frayed cord. A perception system has to make the same call from a camera mounted on a moving robot, with the tool partly hidden by a hand.
That makes condition recognition a different data problem from tool identification, which asks what the tool is. Condition asks whether it should be used, and the answer depends on the tool type, the job and the owner's policy. No public label scheme settles that, so the grading rules are part of what you buy.
What published tool datasets do and do not cover
HANDAL is a public dataset of real-world manipulable objects with category-level 6-DoF pose and scale, handle affordance segmentation and reconstructed meshes, and it focuses on robotics-ready hardware and kitchen tools such as pliers, utensils and screwdrivers (Guo et al., 2023). It is a useful reference for how to capture and annotate tools with a single camera and semi-automated processing.
It does not label wear or damage, and it was not built to. For this page it is only a capture and labeling reference: it shows what pose and affordance annotation can look like, and it tells you which fields to ask a supplier to add. Do not assume a model trained on pose-labeled pristine tools will recognize a worn one. If you also need tool motion, the tool pose tracking specification covers that annotation, and tool selection knowledge covers the textual question of which tool suits which job. Neighboring tasks sit in the maintenance cluster hub.
Who may hold condition-labeled images
| Supplier type | What they may have | Typical gap |
|---|---|---|
| Workshops and repair shops | Photos of failed tools taken for warranty or documentation | Inconsistent framing; condition noted in free text |
| Tool rental and equipment depots | Check-in photos of returned tools, damage notes | Photos target chargeable damage, not graded wear |
| Calibration and tool-service providers | Before-and-after records for instruments and torque tools | Often records rather than images |
| Manufacturers' quality or returns teams | Failure-analysis photos | Possibly under customer or warranty restrictions |
| Training providers | Teaching images of worn versus new tools | Staged, small variety |
Records of this kind are existing assets. They were created to document charges, warranty claims or safety, not to train a model, so expect to normalize text, link images to tool IDs and review what customer information appears in the frame. The related equipment rental inspection and repair page covers the rental cycle in more detail, and the defect image datasets page covers the broader inspection setting.
Designing the label scheme
Ask for the supplier's own condition vocabulary first, then decide whether to map it onto yours. Useful dimensions include:
- Tool type and variant, so a condition label is read against the right expectation.
- Condition grade, with a written definition for each level. Keep it ordinal if your downstream decision is use, repair or discard.
- Damage type: wear, deformation, crack, corrosion, contamination, missing part, loose fitting. Use only the types the supplier can document.
- Location on the tool, such as tip, jaw, handle or cord.
- Decision recorded by the shop: returned to service, repaired, discarded, with the reason if given.
- Image conditions: lighting, background, distance and whether the tool is in hand.
A grade is only as good as its agreement across reviewers. Ask for double-labeled samples and a disagreement log; the dataset documentation practices in Datasheets for Datasets are a reasonable template for the questions a supplier should answer about composition and collection.
Illustrative package
This is a synthetic example, with values chosen only to show structure.
| Field | Illustrative value | Note |
|---|---|---|
| image_id | img_000123 | Pseudonymous |
| tool_type | adjustable_wrench_A | Supplier taxonomy |
| condition_grade | 0 serviceable / 1 worn / 2 unsafe | Definitions supplied in writing |
| damage_types | [deformation, corrosion] | Multi-label |
| region_mask | polygon on jaw | Optional |
| decision | discard | From shop record |
| capture | bench, 45-degree view, diffuse light | Per-image metadata |
| context | in_hand / on_bench | Needed for robot-view realism |
Include both bench-style images, which are clean and consistent, and in-hand images, which resemble what a robot camera would see. A model that has only seen the first will fail on the second.
Limits: what this data cannot support
- No manipulation data. Images tell a model what a worn tool looks like, not how it behaves when gripped, loaded or dropped.
- No safety decisions. A condition classifier gives a prediction. Whether to use, repair or retire a tool is the deployer's policy and, where applicable, the tool maker's guidance; neither is established by an image set.
- Hidden wear. Internal fatigue, calibration drift in a torque tool and electrical faults do not show in a photo.
- Class imbalance. Serviceable tools dominate; severely damaged ones are rare, and they may be disposed of before photographing.
- Label drift across shops. One shop's "worn" is another's "serviceable." Normalize grades before pooling.
- Evaluation leakage. Hold out whole shops or tool models for evaluation and keep them out of training.
Questions to ask suppliers
- How were condition decisions made, and by whom?
- Are written grading definitions available, and have they changed over time?
- Which tool types and brands appear, and how many images exist per type and grade?
- Were images taken under controlled conditions, ad hoc on phones, or both?
- Do images show tools in hand, in storage, or only on a bench?
- Is each image linked to a tool ID, a customer record or a repair outcome?
- What customer names, serial numbers or workplace details appear in frames?
- Were images ever edited, cropped or compressed, and was the original kept?
- Who can license the images: the shop, the rental company or the customer who owned the tool?
Discuss your requirements
If you are looking for tool-condition imagery, describe the tool types, grading scheme and permitted uses to SourceX. SourceX looks for US businesses that hold suitable material and follows up about potential sources; it checks the data and the supplier's licensing permissions, and annotation can be arranged case by case. A request does not guarantee a match. For the overall sourcing map, see the humanoid training data guide, and for a related catalog category see images and inspection photos.
Explore sourcing and licensing options with SourceX
Sources
- HANDAL: A Dataset of Real-World Manipulable Object Categories with Pose Annotations, Affordances, and Reconstructions — tool pose, affordance and mesh annotation reference; does not cover wear or damage.
- Datasheets for Datasets — documentation template covering motivation, composition and collection process.