How Can a Mop Robot Spot Clutter and Alert a Person?

• Mopping robot
• Objects on the floor
• Computer adding up the clutter score
• Phone receiving the alert
• Data path

What you will learn

• Name the signals that show a floor zone is cluttered and merge them into a clutter score per zone.
• Explain what the robot sends to the computer when it uses a camera, and what it keeps back to protect privacy.
• Choose a threshold, a number of confirming passes and the content of an alert, then measure quality with precision and recall.

Before you start

It helps to know the close-range sensors of a mop robot from Which Sensors Stop a Small Robot Falling or Colliding? and how objects are detected in images from How Is Drone Footage Object Detection Different?. This article teaches only the part that is specific to a mop robot.

What is clutter to a mop robot, and what measures it?

Clutter is anything lying on the floor that stops the robot finishing a pass: charging cables, socks, shoes, toys. It blocks the path, gets wet or jams the robot. The robot does not need to see the object to know a zone has a problem, because its own motion leaves traces.

The simplest signal is a bump sensor. A survey of cleaning robots describes robots relying on simple bump sensors to detect obstacles and make route corrections. The illustrative design adds three more signals taken from the robot itself.

For obstacle density the controller can use a costmap such as the one in Nav2. The Nav2 obstacle layer marks obstacles from sensor data. The same documentation describes a denoise layer that filters noise-induced false obstacles from the costmap. That idea matters for clutter: one odd reading is not enough to conclude anything.

• Bumps and jams
• Obstacle density
• Unreachable cell
• Re-planning
• Score-adding step
• High-score zone: alert a person
• Medium-score zone: keep watching
• Low-score zone

What does a camera add, and what does the robot send to the computer?

A camera tells what is lying on the floor, which a bump cannot do. Object detection runs on the computer, not on the robot's microcontroller. The ML Kit documentation describes an API that detects and tracks objects in an image or a live camera feed. It also says the API can classify detected objects with a built-in coarse classifier. Detecting an object does not establish that the API recognizes cables, socks or hazardous floor areas. In the illustrative model, detection is only an additional signal after testing on real floor images; choose a suitable computer-side model rather than treating this mobile API as ready-made PC software.

The robot does not need to stream video. The illustrative design offers two ways to send data:

• A still frame on request: the robot takes one picture after repeated bumps or when the controller asks, and sends that one picture.
• A detection event: if detection runs close to the robot, it sends only the object type, the map position and a confidence.

An object’s image position is not a map position. In the illustrative model, the controller needs robot pose and calibrated geometry to associate it with a floor zone, including uncertainty. Without that conversion, the alert names the observed robot area rather than inventing a precise object coordinate.

This also helps privacy. In a home or office, a floor picture may include people and personal items. NIST's digital identity guidance has a requirement that permits collecting only the personal information necessary for the stated purpose. Applied to the design: process images on a computer inside the building, delete a frame once it has produced a result, keep only the frame attached to an unresolved alert, and tell people in the building that the robot has a camera. This is design reasoning, not something NIST prescribes for mop robots.

• Mopping robot
• One captured frame
• Detection box around the stray object
• Computer inside the building
• Building boundary
• Deleted frame
• Continuous video stream, not sent

How do a score, a threshold and several confirming passes cut false alarms?

The controller adds up the signals of each zone into a clutter score, then alerts only when the score stays high over several passes in a row. Each signal has a weight: repeated bumps in one cell count more than a single touch, and a detected cable or shoe counts more than a shadow. The weights are a design choice to tune in a real building.

Hysteresis uses two levels instead of one. The score must cross the high level to raise an alert and drop below a lower level to clear it. A score hovering around one level then does not make the alert flicker on and off.

Confirmation over several passes complements the threshold. Routine alerts wait for enough above-threshold passes; obstacles are marked and avoided on first contact. A lost safe route or persistent jam requires stopping and an immediate fault alert, without waiting for more passes. The reason for care lies with the person receiving it. A review of alarm fatigue describes a phenomenon in which repeated exposure to frequent or non-actionable alarms causes overload and a gradual loss of sensitivity. The review also says a high frequency of false alarms creates distrust in the alarm system. One measure it lists is adjusting alarm thresholds to manageable levels. For a mop robot, every false alarm makes the reader trust later, real alerts less.

• Score of the zone with the cable
• High threshold to raise an alert
• Low threshold to clear an alert
• Pass on which the alert is sent
• Zone that crosses the threshold for one pass only

What should an alert say, and how is its quality measured?

A good alert says where, why and what to do, in a few words. The illustrative design has four parts: the zone on the map, a still frame if there is one, a short sentence such as "a charging cable blocks the way", and two choices: cleared or snooze. Snoozing makes the zone ignore alerts for a period the person picks, while the robot keeps mopping the rest.

To know whether alerts are good, count two things. Google's Machine Learning course defines precision as the proportion of positive classifications that are actually positive, and recall as the proportion of actual positives classified correctly. For a mop robot, precision tells how many of the alerts sent were for zones that really were cluttered. Recall tells how many of the truly cluttered zones were alerted.

The two pull in opposite directions. Lowering the threshold usually increases recall and can reduce precision; measure both on the same labelled scenarios. The same course says the metric to prioritise when choosing a threshold depends on the costs, benefits and risks of the specific problem. Here a false alarm costs a person's attention, while a missed alert costs a jammed pass or a wet floor where objects lie. Someone has to decide which hurts more.

• Truly cluttered zone that was not alerted (missed)
• Truly cluttered zone that was alerted (correct)
• Alerted zone that was not cluttered (false alarm)

Real-world example

A small office has a corridor zone near the door. This is a design scenario, not a test result. On the first pass the robot bumps into a charging cable lying away from the socket, backs off, marks the cell and goes around. The score crosses the high level for the first time, but there are not yet enough confirming passes for a routine alert. The robot still marks and avoids the cable immediately.

On the second pass the robot jams at the same spot and takes one picture to send to the computer. Detection finds a stray object on the floor, the score crosses the high level for the second pass in a row, and the office manager's phone receives an alert with the map and the picture. The manager clears the cable and taps "cleared". The computer deletes the picture, the zone score drops and the robot mops the skipped part.

In the same session the zone beside a cabinet also scored high because a chair had been pulled out once. By the next pass the chair was back, the score fell below the low level and no alert was sent. Had the system alerted on the first pass, the person would have received a false alarm.

Common misunderstandings

A camera is enough. A camera under a bed gets little light, and pictures exist only when the robot takes one. Bumps, jams and skipped cells still give signals when images are unusable.

The earlier the alert, the better. Early alerts raise false alarms, and frequent false alarms make people lose trust. Waiting one more pass often buys fewer false alarms.

Sending video to the computer is more convenient. The robot only needs to send a frame when asked or a detection event. Continuous video loads the network and adds personal data to keep safe.

Frequently asked questions

How does a mop robot know the floor is messy?

It relies on traces of its own motion: repeated bumps, jams, skipped cells and a path that keeps being re-planned. A camera adds a signal about what the object is.

Does a mop robot need a camera to report clutter?

Not necessarily. Bumps, motor current and skipped cells are enough to mark a zone with a problem. A camera may help distinguish objects once the model is tested on appropriate floor images; without measurements it cannot be claimed to cut false alarms. Images also add a privacy duty.

How can false alarms from a mop robot be reduced?

Use hysteresis, confirm over several passes and let people snooze an alert. Track the precision of the alerts that were sent to tune the threshold.

Summary

• A clutter score per zone merges bumps, jams, skipped cells, re-planning and image detection results.
• The robot sends only a still frame or a detection event, processed inside the building and deleted after use.
• Hysteresis and confirmation over several passes cut false alarms; precision and recall show whether alerts can be trusted.