Short answer
Soft goods such as pouches, bags, sacks and textiles change shape when touched, so their listed dimensions and weights describe only one state. A useful property record adds fill level, stiffness, recovery, surface friction and seam or closure behavior, with the measurement method stated. Listed attributes alone cannot predict handling, and demonstrations are still needed.
The buyer's problem
A rigid carton has one shape. A bag of rice has as many shapes as there are ways to put it down. A catalog page gives its nominal size and net weight, and that is about all. A robot lifting it will find that the contents shift, the film wrinkles, the center of mass moves and the pouch sags between grasp points.
Researchers working on deformables see this as a perception and a manipulation problem at once. The ZeroWaste dataset, for example, was presented as a first in-the-wild, industrial-grade waste detection and segmentation dataset aimed at highly deformable, often translucent objects in extreme clutter on a recycling sorting line, with subsets of 4,503, 6,212 and 1,410 images (Bashkirova et al., 2022). It documents how hard deformable segmentation is; it does not supply handling properties.
This specification describes the property fields a buyer can ask a manufacturer or distributor for, so that images and catalog data come with some physical context.
What this data can support and what it cannot
Property records can support:
- Grouping products by handling class, so a policy or planner treats a stiff-walled pouch differently from a loose sack.
- Choosing grasp strategies to test (pinch, scoop, two-hand support) per class.
- Building test sets that vary stiffness and fill state deliberately.
- Simulation parameters as starting values, to be tuned against real behavior.
- Annotator guidance when labeling deformation states in images.
They cannot support:
- Predicting handling. A stiffness number measured in a lab does not tell a controller what a grasp does. Robot or teleoperation data with contact signals is needed. Published robot datasets such as RH20T record visual, force, audio and action data for each sequence (Fang et al., 2023); a property table has none of that.
- In-situ appearance. Wrinkles, folds and occlusion in real conditions differ from studio photos.
- Aging and environment. Properties change with temperature, humidity and wear.
- Variation between lots. One batch of film is not the next.
Illustrative property specification
The specification below is Illustrative. Field names, units and values are synthetic. Each buyer should select the fields their architecture can use.
# Illustrative only - not an available dataset
record: soft_good_properties_v0
product_ref: catalog_id_or_sku # link to product master data
state_conditions: # properties depend on state
fill_level: [empty, half, full]
temperature_c: 20 # example condition
shape_and_size:
nominal_dims_mm: given_by_catalog
dims_by_state_mm: measured_per_fill_level
mass:
net_mass_g: given_by_catalog
gross_mass_g: measured
center_of_mass_offset_mm: measured_or_null
mechanical:
bending_stiffness: {method: cantilever_test, value: measured, unit: stated}
recovery_after_compression: {method: stated, value: measured}
extensibility: {method: stated, value: measured_or_null}
surface:
friction_vs_steel: {method: stated, value: measured_or_null}
friction_vs_self: {method: stated, value: measured_or_null}
slip_when_wet: {tested: false}
closure_and_seams:
type: [zip, heat_seal, tie, fold]
failure_notes: free_text
measurement:
instrument: stated
n_units_per_sku: stated
lot_ids: listed
Everything depends on state_conditions. A stiffness value without the fill level it was measured at is nearly unusable.
Property families by product type
| Product family | Fields that matter most | Typical blind spot |
|---|---|---|
| Pouches and sachets | Fill level, seal strength, slosh | Contents shift as the pouch moves |
| Sacks and bags | Fill shape, fold stiffness, handle strength | Handles tear before the body does |
| Film and wraps | Cling, tear direction, translucency | Hard to see; hard to grip |
| Textiles | Drape, stretch, thickness | See apparel and textile attributes |
| Foam and padding | Compression set, recovery | Slow recovery misleads quick tests |
| Produce and food packs | Squish tolerance, wet surface | Bruising is not captured by stiffness |
Quality and acceptance criteria
- Method disclosure. Every measured value should name its method and instrument.
- State coverage. Properties should be reported at stated fill and temperature conditions.
- Replication. Ask how many units per SKU were measured and whether the spread is reported.
- Traceability. Records link to product master data and lot IDs.
- Null honesty. Missing values are marked as missing, not zero.
- Image linkage. Photos taken in the same states as measurements, if images are part of the package.
Buyers can set their own thresholds; there is no universal number of units per SKU that is sufficient.
Where the measurements come from
Manufacturers run material tests for packaging and quality control, and some of those tests overlap with the fields above. Distributors rarely hold them. Often the cheapest route is to ask whether existing quality-lab data can be released, then fill gaps with a commissioned measurement step on a sample of units. That second step is a new collection with its own cost, and it would be scoped case by case. For packaging-specific fields see packaging material properties, and for the wider property schema see physical property annotations.
Questions to ask suppliers
- Which of these properties do you test routinely, and with what method?
- Are values reported per fill state, or only for the nominal product?
- How many units and lots were measured per SKU?
- Can internal test reports be released as structured data rather than PDFs?
- Is any test method proprietary or tied to a customer specification?
- How do you handle products whose properties change with temperature or storage time?
- Do you hold images or video of the same units in the tested states?
- Who can license the test data, and is any of it supplied by a materials vendor?
Permissions and provenance
Material specifications can be trade secrets, and some test methods are customer-specific. Brand imagery attached to the records may carry photographer and brand rights. Ask which fields originate with the supplier and which come from vendors, and what uses the license allows. Removing brand names from a record is a preparation step, not a rights solution.
Related reading: material attributes for manipulation for the wider attribute set, product dimensions and weights for the nominal values these records extend, and transparent and reflective objects for film and clear pouches that are hard to image.
Discuss your requirements
If your project handles soft goods, describe the product families, property fields and permitted uses to SourceX. SourceX reviews the requirement, looks for US businesses that may hold suitable property or product data, checks the data and the supplier's licensing permissions, and follows up about potential sources. If no existing data fits, a commissioned measurement could be discussed. A request does not guarantee a match. See also the objects hub, the humanoid training data guide and the catalog category for supply chain and logistics data.
Discuss your humanoid training-data requirements with SourceX
Sources
- ZeroWaste Dataset: Towards Deformable Object Segmentation in Cluttered Scenes — deformable, translucent object segmentation in clutter and subset sizes.
- RH20T: A Comprehensive Robotic Dataset for Learning Diverse Skills in One-Shot — robot sequences with visual, force, audio and action signals.