Site hazard detection module
Lidar Can't See a Wet Floor
An edge VLM that flags spills, warning signs, smoke, and elevator outage notices — distance-gated, de-duplicated alerts that read signage too.
Why floor hazards are hard to detect
Spills have no height
Lidar and most range sensors detect obstacles by height difference. A spill lies flat on the floor, so there's almost nothing for lidar to measure.
Generic object detectors don't know 'wet'
Off-the-shelf detection models are good at edges and shapes, but a spill has no fixed shape or boundary — the model can't tell a dry floor from a wet one.
Slips are one of the most common public liability claims
Slip-and-fall incidents are among the most common public liability claims in hospitals and retail, and no one has the staff to watch every corner of the floor all day.
From camera to resolution
The detection logic is built as a clear chain of responsibility, from capture to the moment staff close the case.
-
1
A camera or robot feed captures the scene
A cleaning, delivery, or patrol robot's onboard camera — or an existing CCTV feed over RTSP — streams continuously to the edge box.
-
2
An edge VLM reads four hazard classes
Cosmos3-Edge runs on the edge and only looks for spills, yellow warning signs, smoke, and elevator outage notices — nothing else.
-
3
A distance gate decides when to fire
A hazard only triggers an alert once it's within a configured distance, so something far across the room or at the frame edge doesn't cause a false alarm.
-
4
The same hazard only alerts once
Once a hazard has been flagged, it won't fire again just because a robot or camera keeps looking at the same spot.
-
5
A structured event goes out
The detection is emitted as a structured event, feeding either the robot's decision layer or the site's existing alerting or ticketing system.
-
6
Site staff respond and close it out
Staff act on the alert — place a warning sign, mop the spill, report the elevator fault — and close the case, completing the chain from camera to resolution.
Deliverables
Edge box image or SDK
An edge-computing box image preloaded with Cosmos3-Edge's detection logic, or an SDK to integrate into existing hardware.
RTSP version for existing cameras
No need to install new hardware on site — plug into an existing CCTV feed over RTSP.
Per-class accuracy report
A measured accuracy and false-alarm report for each hazard class — spills, warning signs, smoke, elevator outage notices — produced once real-footage testing is done.
Prompt and validation workflow for new hazard classes
Adding a new hazard class means adjusting the prompt and running the validation workflow, not retraining a model.
Who it's for
Robot manufacturers
Cleaning, delivery, and patrol robots need a perception module that understands site hazards, but building and training one in-house takes time and data.
License the perception module into an existing product line, and a robot gains the ability to read spills, warning signs, smoke, and elevator outage notices.
Camera and video platform vendors
Platforms want smarter alerting features, but the usual approach trains one model per hazard, which makes adding classes slow.
Ship the detection module as an app — adding a hazard class is a prompt change and a validation run, not a retraining cycle.
Retail, hospitals, and property managers
Slips are one of the most common public liability issues, but no one has the staff to watch every floor all day.
Point an existing camera at the feed and get a structured alert the moment a hazard appears.
Why us
Most comparable vendors train one model per hazard.
One model, multiple hazard classes
We run a VLM for detection — adding a hazard class is a prompt and validation change, not a new model to train.
Reads signage, including Chinese
Notices like '電梯維修中' (elevator under repair) are within scope, not just shapes and colors.
Built to avoid false alarms, not patched afterward
Distance-gated triggering and de-duplicated alerting are part of the detection logic itself, not a filter bolted on after the fact.
The output is a structured event, not a frame
What comes out is a structured event you can feed directly into a robot's decision layer or a site's existing alerting system.
Pricing
OEMs and site operators are priced differently, based on how the module is used.
OEM license
Robot manufacturers license the perception module into an existing product line
- Priced per licensed robot unit
- Model updates and validation for new hazard classes included in the license
Rate depends on shipping volume — to be negotiated
Site alerting service
Retail, hospital, and property teams plug into existing cameras
- Priced per camera or per site, annually
- An RTSP-only option is available without new hardware
Rate depends on site size and camera count — to be negotiated
Add-on: human review
A human review layer on top of automated detection for sites with low tolerance for false alarms
- Automated detection plus a human double-check
- Suited to sites where a false alarm carries a real cost
Review service priced separately
Current status and validation plan
Verified
- On simulated footage, Cosmos3-Edge detects all four hazard classes — spills, yellow warning signs, smoke, elevator outage notices
- Distance-gated triggering and de-duplicated alerting logic is working
In progress
- Still tuning the trigger distance
- Haven't tested real-camera glare, lighting, or viewing angle yet
- No false-negative or false-positive numbers yet
Next
- Record two weeks of real hallway footage on existing cameras to measure accuracy and false-alarm rate for spills and warning signs
- Once the numbers hold up, talk to a video platform vendor about listing it as an app
FAQ
Does video ever leave the site?
Detection is designed to run on the edge, so raw footage doesn't need to be sent to the cloud for analysis; the exact integration depends on the site's setup.
How do you measure the false-alarm rate?
We plan to record two weeks of real hallway footage on existing cameras and measure accuracy and false alarms for each hazard class. We don't have numbers yet — we'll publish them once they hold up.
How long does it take to add a new hazard class?
Because it's a VLM rather than a model trained per hazard, adding a class is a prompt change plus a validation run, not retraining.
Does it read Chinese signage?
Yes. Notices like '電梯維修中' (elevator under repair) are within scope, not just shapes and colors.
Can it work with our existing cameras?
Yes — a planned version connects to existing cameras over RTSP without new hardware.
How is this different from generic object detection?
Generic detectors are good at edges and shapes, but a spill has no fixed shape and lidar can't see its height difference. We target four specific hazard classes — spills, warning signs, smoke, elevator outage notices — and read the accompanying signage too.
Let's talk
Whether you're a robot manufacturer, a video platform vendor, or a site operator, tell us what you need and we'll follow up based on your role.