Short answer
Handling-induced damage examples show harm done to an item by the way it was gripped, moved or set down, such as grasp marks, scuffs and dents. Manufacturers and 3PLs may hold images of such damage. Images show the result, not the handling that caused it, so teaching a robot gentle handling needs robot, teleoperation or instrumented-contact data.
The problem robots add
Human handlers learn what damages a product through feel and experience: a finish that marks under a gloved thumb, a carton that crushes at the corners, a panel that flexes. A robot gripper applies force differently and often with more repeatability, so it can create the same mark on every unit. Buyers building manipulation systems want examples of what "damaged by handling" looks like, and ideally what handling caused it.
This is a narrower question than general damage. Transit damage from drops and crushing is covered in shipping package damage. Damage from the robot itself, when a policy fails, belongs with robot manipulation failure datasets. This page is about marks and deformation left by grasping, clamping, stacking, sliding and placing, whoever or whatever did the handling.
For the cluster overview, see the inspection, defect and failure data hub. The whole section is mapped in the humanoid training data guide.
Where existing examples sit
Existing material is scattered across routine records:
- Quality rejects in manufacturing, where finished parts show clamp or fixture marks, scuffs from conveyors, or dents from bins.
- Receiving and put-away exceptions in 3PLs, where inbound or outbound items are photographed because of damage claims. See 3PL providers as data suppliers.
- Returned-item grading, where condition notes describe marks; see returned item condition grading.
- Incident and near-miss records with photos; see near-miss and incident reports.
The record usually states "damaged" and sometimes "caused by forklift" or "crushed in stack." It rarely says which handling step created a mark, how hard the grip was, or what the tool was.
A reference for capture, not for the topic
ARMBench was collected in an Amazon warehouse pick-and-place workcell and contains images, videos and metadata for more than 235,000 pick-and-place activities on more than 190,000 unique objects, with tasks including object segmentation in clutter, object identification and defect detection (Mitash et al., 2023). Its scope is pick-and-place in one type of workcell. It is only a capture and metadata reference here: it shows how images and videos of handling events can be paired with per-activity metadata. Do not read it as a source of handling-induced damage labels, and do not assume it establishes what damaged items look like in your setting.
What existing damage images can support
- Appearance models that recognize scuffs, dents and grasp marks on your product family.
- Acceptance definitions, if images are paired with a disposition (shipped, discounted, scrapped) so you can see where the line sits. See failure taxonomies for embodied tasks.
- Post-handling checks, such as a robot inspecting an item after placing it.
- Evaluation of whether a handling policy damaged items, when you hold out items by SKU or finish.
Limits: what this data cannot support
It cannot teach gentle handling. A photo of a mark does not tell the model the grip force, contact area, approach speed, surface condition or tool that caused it. A policy needs closed-loop robot or teleoperation data, and often force or tactile sensing, to learn what to avoid. See robot manipulation failure datasets.
It cannot show causes without extra context. Many marks come from several events. A scratch noticed at receiving may have been made at the supplier, in transit or on a shelf.
Existing images are biased toward damage noticed and photographed, often for claims. Mild marks that nobody photographed are absent. Lighting and angle are inconsistent.
Commissioned collection as a design option
Where existing records lack causes, a buyer can consider a purpose-built capture: items handled by a known gripper or fixture under scripted conditions, then photographed under controlled light before and after. This allows before-and-after pairs, known handling parameters and consistent viewpoints. It costs time and money, needs site access and still requires consent for any people in frame. The trade-offs are covered in existing assets, new collection or specialist partners. It can capture robot or handheld-device action only if such equipment is used, and contact forces only if the rig is instrumented.
Illustrative record design
The following is an illustrative design with synthetic values, not an existing dataset.
| Field | Example | Notes |
|---|---|---|
| item_id, sku_group | item_0231, painted_panel | Hold out by SKU group |
| before_image, after_image | img_a_0231, img_b_0231 | Same viewpoint and light |
| handling_event | clamp_transfer | Free text is weaker than a controlled list |
| handling_params | gripper_type, pad_material | Only if known; otherwise "unknown" |
| damage_label | grasp_mark / scuff / dent / none | Includes "none" |
| severity | minor / major / reject | Defined by the buyer's acceptance rules |
| disposition | shipped / scrapped | Optional |
Quality and acceptance criteria
Set your own thresholds, then test on a sample: a reviewer agrees with the label in most cases; before-and-after pairs align closely enough to compare; "none" examples exist for each SKU group; severity follows a written rule rather than individual judgment. Check agreement between reviewers as in inter-inspector agreement.
Permissions and provenance
Damage records can implicate named carriers, customers, vendors or employees. Photos may show product designs, labels or workers. Ask who owns the images, whether claims or legal disputes are attached, and whether customers or suppliers must approve release. This is not legal advice.
Questions to ask suppliers
- Do records identify the handling step that caused the damage, or only that damage exists?
- Are before-handling images available for the same items?
- How were damage labels assigned, and by whom?
- Are undamaged control items photographed under the same conditions?
- Which images were taken for claims, and are any under dispute?
- What lighting, devices and viewpoints were used?
- Are items unique, or are many identical SKUs included?
- Could a controlled handling capture be arranged at your site, and what consent would be needed?
- Who can approve licensing of images that show customer products?
Discuss your requirements
If you need existing damage images, or want to discuss a controlled capture, describe the product, handling steps 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. If no existing asset fits, SourceX can scope a commissioned collection of new recordings; a request does not guarantee a match. The physical-world collection page describes the separate path for new recordings.
Discuss your humanoid training-data requirements with SourceX
Sources
- ARMBench: An Object-centric Benchmark Dataset for Robotic Manipulation — warehouse pick-and-place images, videos and metadata, as a capture reference only.