Short answer
Sign and label recognition data consists of images of safety signs, floor markings and equipment warnings, labeled by meaning, ideally in real facilities under varied light and angles. ISO 7010 gives a reference set of registered safety signs to define classes. Recognizing a sign is not the same as obeying it, so buyers still need rules, maps and safety validation.
Why signs matter to a robot that works near people
Workplaces tell people what to do with signs: do not enter, wear eye protection, forklift crossing, hot surface, emergency exit. They also mark constraints that no part of the robot's own sensing would reveal, such as a restricted zone or a valve that must stay closed. A humanoid that reads these cues can ask for help, avoid an area or flag a missing sign. A humanoid that cannot read them behaves as if the constraints do not exist.
The data need is narrow and well defined. Signs are standardized symbols on high-contrast backgrounds, which makes detection feasible, but real examples are faded, dirty, partly hidden, mounted at odd heights and glossy under overhead lights. This guide covers where such images come from, how to define the classes, and what to demand from a supplier.
Defining classes with ISO 7010
ISO 7010 prescribes safety signs for accident prevention, fire protection, health hazard information and emergency evacuation. Their shapes and colors follow ISO 3864-1 and symbol design follows ISO 3864-3; the standard excludes rail, road, river, maritime and air traffic signalling. It has been amended repeatedly, with Amendment 11 published in March 2026, and a replacement draft is at enquiry stage (ISO 7010:2019).
For a buyer the useful consequence is a stable class vocabulary: each registered sign has an identifier and a meaning, so a label can be "ISO sign E001 (emergency exit)" rather than a free-text caption. Because the standard changes, record the edition and amendment against which labels were assigned. Many facilities also use signs that are not in ISO 7010: company-made notices, older ANSI-style signs, handwritten warnings and temporary tape. A dataset limited to registered symbols will miss them, so decide whether your classes include a catch-all category such as "other text sign" and how its text is read.
A related set of labels, chemical hazard pictograms on containers, follows a different regime; see hazmat and GHS labels. The US hazard communication rule, 29 CFR 1910.1200, covers container labels, safety data sheets and worker training. Do not treat sign images as a source of compliance knowledge.
What this data can support, and what it cannot
It can support detection and classification of sign types, reading of text on signs with OCR, localization of a sign in a scene (which door, which machine), and evaluation of how recognition degrades with distance, angle and damage.
It cannot tell a robot what to do. A sign reading "authorized personnel only" has a meaning that depends on who the robot is, who authorized it, and which policy its operator set. Rules for responding belong in the control and safety layer and are validated separately; data from images does not establish safe behavior. Sign recognition also cannot detect what is missing: absence of a sign is not evidence of safety. And images do not carry facility maps, so the link from "wet floor sign" to a hazard region needs scene understanding or a site map.
Sources of sign images
| Source | Strengths | Weaknesses |
|---|---|---|
| Sign manufacturers and distributors | Clean reference images across the whole range | Not in situ; commercial image rights |
| Facility safety audits and inspection photos | Real placement, wear and context | Photographs include people, screens and unrelated third-party items |
| Facilities and EHS teams, by commissioned walk-through | Designed coverage of angles and distances | Requires site access and agreement |
| Public image collections | Wide variety | Mostly roads and general scenes; rights vary |
| Synthetic renders | Controllable variation | Need real validation |
Safety audit photos deserve particular care. They exist because someone documented a problem, so they over-represent damaged or missing signs, which is useful for edge cases and misleading as a base rate. See images and inspection photos for how this asset type is described on the site.
Illustrative label specification
The block below is Illustrative; field names and examples are synthetic.
# Illustrative sign-image annotation record - synthetic values
image_id: img_000123
site_id: site_017
camera: {height_m: 1.6, type: head_mounted_rgb}
objects:
- bbox: [412, 188, 96, 96]
sign_class: iso7010_E001 # registered sign id when it applies
sign_edition: "ISO 7010:2019 + Amd 11"
text_content: "" # transcribed if text is present
non_standard: false
condition: [faded, partially_occluded]
mounting: wall_high
viewing_angle_deg: 35
distance_m_estimate: 6
scene_relation: door_id_D4 # optional link to a site map entry
privacy: faces_blurred: true, screens_masked: true
license_id: lic_site_017
Quality and acceptance criteria
- Class definition. The supplier states the class list, the standard edition and the rule for unlisted signs.
- Condition coverage. Counts by condition (clean, faded, damaged, occluded), by distance and by mounting height, so you can see whether far, small signs are included. Resolution at the sign matters more than image size.
- Hard negatives. Images with look-alike objects: red-and-white stripes, posters, reflective tape, general notices.
- Text accuracy. Where text is transcribed, compare a sample to the image by hand.
- Site diversity. Counts by site and sign vendor; one facility's signage style can dominate.
- Held-out sites. Keep entire facilities out of training for evaluation.
Permissions and provenance
Facility photos often capture people, badges, screens and equipment of third parties. Ask whether the people in the images were informed and whether any consent covers training use; see worker consent and identifiable people. Blurring faces does not by itself settle permission questions. Sign artwork may be vendor-owned, and a customer's site may restrict photography for security reasons. This is a diligence list, not legal advice.
Camera viewpoint and distance
Signs are designed for people who look at them from a few meters at eye level. A humanoid's cameras may sit at a similar height, but a wheeled base, a crouching pose or a body-mounted fixed camera changes the angle. State the intended camera height range and the farthest distance at which the robot must read a sign, then ask for images that cover that range. Resolution on the sign is the binding constraint: a small warning label that is legible at one meter may be a few pixels at five.
Also consider illumination: reflective and photoluminescent signs, such as exit signs in a dim corridor, change appearance sharply with lighting and viewing angle. A dataset taken only in bright daytime scenes will not cover them.
Limitations and extra preparation
Sign data supports perception only; object-level cautions are covered in hazard properties of objects. You still need a rule base mapping each sign class to an action, a site map, and validation of the whole behavior under your safety process. Signs change as sites change, so refresh. Related perception topics include equipment controls and interfaces, PPE recognition and text on packages and labels.
Questions to ask suppliers
- How many distinct sign types appear, and which standard edition defines them?
- Were images taken at working distances and heights for a robot, or at eye level up close?
- What share shows damaged, faded or partly hidden signs?
- Did the photographers have site permission, and are people in the images covered by notice or consent?
- Are custom, handwritten and temporary notices included and labeled?
- Can you provide site and camera metadata so that we can split by facility?
- Who owns the sign artwork shown in the images, and does the site restrict reuse?
- How often are sign inventories refreshed?
Discuss your requirements
Define the sign classes and the facility types first. When you describe the sign and label requirement to SourceX, SourceX reviews it and follows up about potential sources among US businesses and facilities. A request does not guarantee a match, and nothing is contracted until a supplier agrees. For wider context, see the humanoid training data guide and the object data hub.
Describe the workplace demonstration data your robotics project needs
Sources
- ISO 7010:2019 Graphical symbols - Safety colours and safety signs - Registered safety signs — scope, structure and amendment status of registered safety signs.
- 29 CFR 1910.1200 - Hazard Communication — US rule on container labels, safety data sheets and training.