Short answer
Environmental hazard images show conditions in a facility that need action: liquid on floors, blocked aisles, fallen items, broken fixtures and leaks. Sources are facilities teams, cleaning contractors and incident files, supplemented by staged capture. Value depends on viewpoint match to the robot's cameras, honest negatives and clear labels; they train recognition and triage, not cleanup or avoidance behavior.
The detection problem
A robot or camera that moves through a building must notice what is wrong with the space, not only with a product. A clear puddle on polished tile, a pallet left in a walkway, a door propped against a fire exit, a cracked glass panel or a drip from a ceiling tile are all hazards, and none looks like the objects most perception datasets contain.
Treat these as potentially hard visual cases and check how a sample handles them: transparent liquids may barely contrast with the floor, and reflections may create false positives. Obstructions are defined by context: a box is fine on a shelf and a hazard in an aisle. So the images need scene context and a clear definition of what counts.
Hazards in the environment are different from defects in products, covered by defect image datasets, and from failures of a task in progress, covered by service-task failure examples. This page covers still-image hazards in the facility itself.
Hazard categories
| Category | Examples | Detection difficulty |
|---|---|---|
| Liquid on floor | Spills, leaks, condensation, mop water | Transparent, reflective surfaces |
| Obstruction | Pallets, carts, boxes, cords in walkways | Needs context on where objects belong |
| Fallen or misplaced items | Dropped tools, stock, trash | Variation in size and clutter |
| Damaged fixtures | Cracked glass, broken tiles, loose rails | Fine detail, lighting dependent |
| Egress and access | Blocked exits, propped doors | Rules vary by facility |
| Surface condition | Ice, debris, uneven flooring | Seasonal and location dependent |
Equipment-level wear and damage have their own page: equipment wear and damage images. Debris and objects that should not be present are treated in foreign object detection, and structural findings in building and infrastructure defects.
Who holds this imagery
Facilities and property managers. Work orders with photos of leaks, damaged fixtures and cleanup requests. Check whether photos are phone quality and taken after the fact.
Cleaning and janitorial contractors. Before and after photos for quality assurance, spill responses and incident logs. Ask whether these show the hazard and the cleaned state of the same place, which would be useful for negatives.
Safety and risk teams. Incident and near-miss files, with images attached to some entries. Rich in context, sensitive in content; see near-miss and incident reports.
Fixed-camera operators. Security or operations video where hazards appear incidentally. Check viewpoint height, resolution and whether people are in frame.
Round-based inspectors. Walkthrough photos from rounds; see operator rounds and walkdowns.
Existing images or staged capture
Existing images are real but may be uneven. Check whether they skew to hazards someone thought worth photographing and to phone viewpoints at human eye level, which may differ greatly from a humanoid's head camera or a mobile base's low camera. Hazards that were cleaned up before anyone photographed them will be absent.
Staged capture lets the buyer set the camera height and path, vary the surface and liquid, and capture before, during and after states. It is limited by what staff will safely stage and by site access. Ask whether facilities teams or cleaning contractors would stage spills of water or approved test liquids on defined surfaces, move carts into aisles and place items on floors. Staged scenes cover what someone designed; they miss the odd cases real buildings produce. Mixed collections are often a reasonable plan: existing images to learn variety, staged images to cover viewpoints and negatives. Describing the requirement to SourceX is the starting point for either route.
What the data can and cannot support
It can support hazard detection and segmentation, severity triage (for example wet floor versus clear), change detection against a clean reference, and evaluation of a perception module's false alarm rate when negatives are included.
It cannot support:
- Cleanup or avoidance behavior. Knowing there is a spill does not teach a robot to mop it or route around it; that needs robot data and a controller.
- Physical properties. A photo does not say whether the liquid is slippery, hot or corrosive.
- Rules about what counts as a hazard. Facility policies differ, and labels encode one facility's view.
- Dynamic events. A still image shows a state, not the moment something fell.
Illustrative package
| Component | Illustrative content |
|---|---|
| Images | Hazard, near-hazard and clean views of the same places |
| Viewpoint tags | Capture height, tilt, device type |
| Scene context | Area type (aisle, kitchen, lobby, loading dock), floor material, lighting |
| Labels | Category, extent mask or box, severity, action required |
| Pairing | Before and after images with a shared location ID |
| Negatives | Reflections, wet-looking but dry floors, stored items in correct places |
| Time metadata | Date, shift, indoor or outdoor |
| Privacy fields | People visible flag, signage visible flag |
Ask for hard negatives explicitly. A set made only of obvious hazards teaches a model to alarm on glare.
Acceptance pointers
- Compare capture heights and angles with your robot's camera geometry.
- Check that transparent liquids are labeled with careful extent, ideally verified by a second view.
- Review a random sample for label consistency; the same pooled water should not be a hazard in one image and not in another. See inter-inspector agreement and label noise for agreement methods.
- Hold out facilities, not only images, for evaluation.
- Confirm the proportion of negatives and near-negatives.
Limitations and permissions
Facility images show people, brands, signage and security details. Cleaning contractors may be recording a client's building under a contract that restricts use. Photos of employees at work, even incidentally, raise consent and notice questions; see worker consent and identifiable people in workplace recordings. Blurring faces does not by itself settle whether the images can be used for training.
Staged capture involves real people on real floors. Cleaning and facility work already carries ergonomic exposure; OSHA describes forceful exertion, repetitive motion and awkward postures as risk factors for musculoskeletal disorders (OSHA). A capture plan should avoid adding strain or slip risk, keep staged hazards controlled and marked, and be reviewed with the site's safety lead.
Matching images to the robot's view
A frequent mismatch is viewpoint. Phone photos taken at standing height, looking down at a spill, differ from the images a robot with a head camera at a different height, or a low mobile base camera, will see. Specify the camera heights, tilt and field of view you expect, and ask the supplier to provide images at those geometries or to describe the geometry of what they hold.
Lighting also matters. Check whether spills in a sample are visible mainly as specular reflections of overhead lights; if so, the same spill may vanish at a different angle. Request images of the same hazard from several angles, and ask for different times of day if windows or skylights change the reflections. A set that shows each hazard from a single angle teaches the model that angle.
Consider multimodal cues. Depth or thermal images can help with obstructions and some leaks, but they exist only where the supplier recorded them. If your platform uses such sensors, say so in the requirement. Facility-maintenance context is covered in facility maintenance tasks, which describes what staff do once a hazard is found.
Questions to ask suppliers
- What device and capture height were used, and can you show sample images?
- Are images of hazards matched with clean or resolved states of the same location?
- What proportion of the images are negatives or near-negatives?
- Who labeled hazards, and by what written definition?
- How many facilities, floor types and lighting conditions are covered?
- Do the images show people, brand marks or security-sensitive areas, and can they be licensed with those?
- Does your client contract allow use of images taken in their building?
- Can staged capture be arranged at your sites, with what safety controls?
Discuss your requirements
If hazard recognition is a gap in your facility perception stack, describe the hazard classes, viewpoints and permitted uses to SourceX. SourceX reviews the requirement and follows up about potential sources, 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. Start from the humanoid training data guide or the inspection, defect and failure data hub, and see the catalog page on images and inspection photos.
Describe the workplace demonstration data your robotics project needs
Sources
- Ergonomics - Overview — ergonomic risk factors and musculoskeletal disorders relevant to cleaning and staged-capture planning.