Short answer
Hazard properties are structured attributes on an object record: sharp edge, hot surface, heavy or awkward load, pressurised or stored energy, hazardous contents. Each attribute needs a named source of truth, such as a manufacturer declaration, a safety data sheet or a measurement. They inform handling rules and evaluation, but do not by themselves make a robot behavior safe.
Why hazard attributes are their own data problem
A humanoid that picks up a ceramic mug and one that picks up a heated skillet should behave differently, yet both are "cup-like objects" to a recognition model. The difference lives in properties that images show poorly or not at all: temperature, mass distribution, edge geometry, internal pressure, contents. Teams that want a robot to modulate behavior need those properties as data attached to the object, either at recognition time (look it up from a record) or at perception time (estimate it from sensors).
This page specifies the first kind: hazard attributes attached to object records. It builds on the general object attribute schema and on physical property annotations, and it is aimed at evaluation leads who need to test whether a system handles flagged objects differently.
What each hazard class needs
Hazard classes differ in how knowable they are from records.
| Hazard | Typical record source | Can an image show it? | Main data gap |
|---|---|---|---|
| Sharp edge or point | Product type, manufacturer description | Sometimes (blade visible), not when sheathed | Edge location and which side is safe to hold |
| Hot surface | Equipment spec, operating state | Rarely; state changes over time | Current temperature; cool-down time |
| Heavy or awkward load | Declared weight, dimensions | Size only | Actual weight, center of mass, grip points |
| Pressurised or stored energy | Product category, rating plate | Rarely | Current pressure, spring or battery charge |
| Hazardous contents | Container label, safety data sheet | Label text, if legible | Contents of unlabeled or decanted containers |
| Fragile or breakable | Packaging, manufacturer note | Sometimes | Threshold force; material state |
| Pinch or crush points | Equipment design | Yes for visible mechanisms | Behavior of moving parts |
Declared and measured values are not the same. A catalog weight is a nominal value; a real unit varies and may be filled or damaged. For a measurement-standard view of nominal attributes, see product dimensions and weights.
Using existing hazard frameworks
Two published frameworks give vocabulary and structure without turning your dataset into compliance evidence.
For chemicals, OSHA's Hazard Communication standard requires manufacturers and importers to classify chemicals and pass the information on through container labels, safety data sheets and training; it was aligned with the UN Globally Harmonized System in 2012 (OSHA, 29 CFR 1910.1200). A hazardous-contents attribute on an object record can point to the safety data sheet and label elements. How labels are recognized in images is covered in hazmat and GHS labels.
For heavy loads, the NIOSH Revised Lifting Equation estimates the risk of low back disorders in manual lifting using horizontal, vertical, distance, asymmetry, frequency and coupling factors and produces a lifting index (NIOSH, 2021). It was designed for human workers, not robots, so it does not give a robot's load limit. It is still useful to a data specification because it shows which object and task attributes matter for a load: where the hand holds it, how far from the body, how often, how good the grip is. Those become fields: handle presence, grip quality, mass, size.
What hazard attributes can and cannot support
They can support filtering and gating (do not auto-handle objects flagged corrosive), evaluation sets stratified by hazard (does the system treat flagged items differently), language grounding of cautions ("careful, it's hot"), and risk-aware planning inputs.
They cannot establish that a behavior is safe. A label "sharp" does not say how to hand a knife to a person; a mass value does not tell the robot whether its actuators can lift it. Hazard attributes are inputs to a safety design; they are not a substitute for risk assessment, testing and a safety case for the specific robot and site. Missing attributes also need handling: absence of a hazard flag must mean "unknown", never "safe".
Illustrative attribute record
The record below is Illustrative; fields and values are synthetic.
# Illustrative hazard attributes - synthetic values
item_id: item_00931
name: stock pot, 20 qt
hazards:
- type: heavy_load
value_source: measured
detail: {empty_mass_kg: 3.1, full_mass_kg: 22.0, grip_points: [two_side_handles]}
- type: hot_surface
value_source: state_dependent
detail: {state_signal: required_from_sensors_or_site_system}
- type: hazardous_contents
value_source: not_applicable
confidence: {mass: measured, hot_surface: category_inference}
unknown_fields: [center_of_mass, handle_temperature_limit]
evidence: [spec_sheet_ref_12, annotator_note_44]
annotator: annotator_A12
reviewed_by: reviewer_B03
license_id: lic_000
The value_source field is the key design choice: categories could be manufacturer_declared, safety_data_sheet, measured, category_inference and annotator_judgment. They let an evaluator weigh a measured mass differently from an inferred one, and they stop an annotator guess being treated as a fact.
Acceptance tests
- Source coverage. For each hazard attribute, the share of records by value source. A dataset made entirely of category inference is a taxonomy, not measurement.
- Double annotation. Two reviewers label a sample independently; report agreement by hazard class and resolve disagreements by rule.
- Unknown handling. Check that unknown is represented explicitly.
- Negatives. Include clearly benign look-alikes (a plastic knife, a cold pan) so a model cannot learn a shape shortcut.
- Link integrity. Each hazardous-contents flag resolves to an actual label or sheet.
- Hold-out. Keep some hazard categories and some sites out of training for evaluation.
Permissions and provenance
Safety data sheets and manufacturer specifications are authored by the producer, who may restrict reuse; facility hazard registers can reveal process information a business treats as confidential. Ask who owns each source document and whether annotation notes may be redistributed. If a record is derived from a regulator's text, check what the license allows. This is not legal advice.
Who should annotate
Hazard labels are judgment calls on top of facts. A generalist annotator can copy a declared mass from a specification, but deciding whether a utensil counts as sharp enough to flag, or whether an awkward load needs two hands, takes domain knowledge. Specify the rubric with examples for each hazard class and borderline cases, and have domain reviewers audit a sample. Record disagreements instead of forcing a vote; borderline items are the most informative for evaluation, because they show where reasonable people disagree about what the robot should do.
Limitations and extra preparation
Hazard state changes; a flag set at record time may be wrong at handling time, so the robot needs live sensing for temperature and contents. Annotation of hazards benefits from domain experts, such as chefs for kitchens or technicians for tools, which raises cost. Packaging can hide or change a hazard, so read packaging material properties alongside. Pair this specification with signage recognition (workplace signage and safety labels) and material attributes (material attributes for manipulation).
Questions to ask suppliers
- Which hazard attributes are declared by manufacturers, and which were assigned by your staff?
- Can you link each attribute to an evidence document, and may we receive those documents?
- Are masses measured or nominal, and for how many records?
- How do you represent unknown, as opposed to not hazardous?
- Who annotates hazards, with what training, and is there a second review?
- Are decanted or relabeled containers included, and how were their contents determined?
- Which source documents carry third-party restrictions?
- How are changes to a product's hazard status after a redesign recorded?
Discuss your requirements
State the hazard classes your system must treat differently and the evidence you would accept. When you describe the hazard-attribute requirement to SourceX, SourceX reviews it and follows up about potential sources; annotation can be arranged case by case, and a request does not guarantee a match. Start from the humanoid training data guide or the object data hub.
Tell SourceX what your evaluation set needs to cover
Sources
- 29 CFR 1910.1200 - Hazard Communication — chemical classification, labels, safety data sheets and training.
- Applications Manual for the Revised NIOSH Lifting Equation (DHHS (NIOSH) Publication No. 94-110) — factors that determine lifting risk for human workers.