A cleaning robot is bought to remove a variable — the person who calls in sick. Then it acquires a new variable of its own: it degrades. Not dramatically, and not on one bad night, but by a few per cent a month in three places at once. By the time it stops mid-aisle, the trend has been visible for weeks in data nobody collected.
1. Two Shapes of Failure
Sudden faults have an event. A collision, a fall from a ramp, water ingress at a wash-down bay, a pallet left in the aisle at 02:00. These are managed with response time: someone is called, someone arrives, the clock is measured.
Gradual faults have no event, only a trend. Capacity slips a season at a time. The map falls behind a rack that moved in March. Dust builds on a lidar window until the robot hesitates at a corner it used to take cleanly. Nobody calls, because nothing has failed yet — so response time, however good, never starts.
| Property | Sudden fault | Gradual fault |
| Trigger | An event: impact, water, obstruction | No event; accumulated wear and drift |
| Detection | Immediate, visible | Weeks later, as a complaint |
| Useful response | Response time, spare on site | Records dense enough to show the trend |
| Typical evidence kept | Repair ticket with date | Usually nothing |
| Cost shape | One repair, one lost shift | Slow loss of cleaned area, then a large bill |
The record that matters: most sites keep only the repair ticket, which carries the date of failure and nothing else. It cannot show a trend, because a trend needs the readings taken while the machine was still working.
2. Battery: Capacity Fade Is Physics
"It no longer finishes a shift" has three unrelated causes, and only one of them is the pack wearing out. Treating all three as "replace the battery" is how sites spend money and stay slow.
Capacity fade is physics, and three habits accelerate it: running close to empty as a matter of routine, charging a hot machine in a closed corner straight after a shift, and storing at full charge across a long holiday. None of these breaks a rule. All of them cost cycles — and the cost is invisible until the shift that used to end at 06:00 starts ending at 04:40.
| Cause | What the operator sees | What to check |
| Capacity fade | Shift ends short, month after month | Area per shift trend; capacity test |
| Contact film at the dock | Charge takes longer; docks early with a flat pack | Charge current; contact condition |
| State-of-charge drift | Docks at 30%, or runs to 10% before docking | Calibration history |
Replace on measured behaviour rather than on years in service: effective area per shift, returns per shift, and time from low-battery warning to full charge. Two months of movement in one direction is enough to justify a capacity test. Age alone is a poor criterion, because two packs of the same age in two sites rarely behave alike.
Stop-work condition: a swollen, deformed, leaking or hot pack is not a maintenance item. Isolate the machine, stop charging, and follow the manufacturer's handling instructions and the site's rules for damaged batteries.
3. Two Cheap Causes: Contacts and Drift
Contact film is the cheapest cause and the one most often missed. Industrial air carries fines; car parks carry oil mist. After a few hundred dockings, the charging contacts wear a film that does not stop the charge — it lowers the current. Four hours becomes six. Day shift sees nothing wrong, because the machine is always full when they arrive. Night shift sees a robot docking early with a flat pack and blames the battery.
Cleaning is a two-minute job with rules: a dry lint-free cloth, contact cleaner used sparingly, never scrape with metal, never rinse with water. Then look at the dock itself — a worn or misaligned contact plate produces the same symptom as a dirty one.
State-of-charge drift is third. The percentage on screen is estimated from voltage and current, and after months of shallow cycling the estimate wanders. Thirty per cent and it docks; or ten and it still finishes. One calibration per the manual resets the estimate — once, when the symptom appears, not monthly. Calibrating on a schedule costs a cycle each time and teaches nothing.
4. The Building Moved, Not the Map
"Localisation is off" usually means the environment changed and the map never learned. Five changes cover most cases, and none of them is a robot fault.
| What changed | Symptom the operator reports |
| Racks shifted, added or removed | Stops at the same rack every lap |
| Pallets left in the main aisle | Repeated detours; one strip never cleaned |
| Floor re-coated or re-marked | Edge distance grows week by week |
| Reflective or transparent surfaces added | Phantom obstacles near glass and polished steel |
| Doors, curtains or ramps appear and disappear | Failed docking; route abandoned at one point |
Walk the floor before touching the robot. Nine times in ten the answer is a pallet, a moved rack or a new mirror finish, and no amount of rescanning fixes a map that was correct for a building that is no longer there.
Then treat the map as an asset rather than a setting. Keep a change sheet in the plant room: date, location, what changed, who logged it. Rescan in a fixed weekly window with the aisle empty — scanning during loading turns temporary obstacles into permanent walls. Version the file by date, keep the previous one, run both for a week before deleting either, and limit editing rights to two named people. A map is not a file; it is the robot's understanding of your building, and it deserves the same control as a drawing.
5. Dust on a Lens Is Distortion
Perception sits in four layers: lidar for the outline of the space, depth or vision for low obstacles, cliff and ultrasonic sensors for steps and glass, and bumpers as the last physical line. The first three have windows, and windows collect dust, oil mist, detergent thrown up by the scrubber, and scratches from the wrong cloth.
| Layer | What it protects | Failure signature | Cleaning interval |
| Lidar window | Walls, racks, route outline | Stops at phantom obstacles; drifts from walls | Every shift in dust; daily in commercial space |
| Depth / vision | Low obstacles, pallets, feet | Late braking; contact with low objects | Daily; after each run in oil mist |
| Cliff / ultrasonic | Steps, ramps, glass | False cliff stops on dark flooring | Weekly, and after any floor re-coat |
| Bumpers | Everything the sensors missed | Physical contact, scuffed covers | Inspect weekly; replace when cracked |
The signature of a dirty window is intermittency. Phantom obstacles. Unplanned stops on a route that worked yesterday. Growing distance from walls. False cliff detection on dark flooring. Failed docking. Because the film is uneven and the light changes through the day, the fault comes and goes — which is exactly why it gets dismissed as "the robot being weird" for three weeks.
Clean dry, with a lint-free cloth or lens tissue, from the centre outward, in one direction, never in circles. Blow grit off before touching the surface; wiping grit is sanding the window. No pressure washer, and no solvent that fogs plastic. Log the cleaning — cleaned and provably cleaned are different claims when a fault is later disputed.
Do not "fix" a dirty sensor by replacing it. Clean, then check calibration and firmware. A fault that comes and goes is an optical or environmental symptom far more often than a failed component.
6. Three Signals and One Card
Three numbers, collected on the same route and the same time window, will show degradation before it becomes downtime.
Effective area per shift. Square metres actually cleaned, pauses excluded. Flat means healthy; a slow decline means capacity, contact film or navigation losses.
Returns per unit area. How often the robot goes back to the dock for a given area. Rises when the pack weakens or the charger under-delivers.
Alarm-to-recovery time. Split into self-recovery and intervention. This is the only one of the three that measures operations rather than repair speed.
Write these into the maintenance sheet, one line per shift:
Area cleaned and route completed, with pauses excluded
Number of returns to dock, and the state of charge at each return
Alarms raised, and whether the robot cleared them alone
Time from low-battery warning to full charge
7. The Monthly Card: Eight Ticks
Half an hour a month, one person, one card. The point is not the inspection; it is that the card exists in a folder when the robot stops and someone asks how long this has been going on.
| # | Check | Pass condition |
| 1 | Battery condition and charge duration | No swelling; charge time within 10% of baseline |
| 2 | Dock and charging contacts | Clean, aligned, no pitting or film |
| 3 | Lidar window and camera lenses | Clear when viewed at an angle in light |
| 4 | Cliff and ultrasonic sensors | No false stops on a known dark patch |
| 5 | Brushes and squeegee | Tufts full; lip flat and contacting evenly |
| 6 | Filters and tank seals | Airflow normal; no residue trail at seals |
| 7 | Map version against a floor walk | No rack, door or coating change unlogged |
| 8 | Firmware and log export | Current version; logs archived off the machine |
Eight ticks means inspected. Anything else goes on a dated action list with a name against it — an unticked box with no owner is a note, not a control.
8. Alarm Triage in Three Levels
Alarms do not deserve equal urgency, and treating them equally is why night calls get ignored by month three.
| Level | Examples | Response |
| 1 | Safety stop; robot stopped in an aisle; bumper event with a person nearby | Attend immediately, on shift; escalate to the named duty person |
| 2 | One sensor offline; repeated docking failure; navigation lost | Fix within the shift; never leave overnight |
| 3 | Consumable warnings; low battery; full tank | Schedule into the next planned service window |
Then name who attends, and put the response target on the wall where the shift can see it. A camera feed into a control room shows that the robot is stopped; it does not move the pallet or wipe the oil film. Remote viewing is information, not attendance — and contracts that confuse the two are the ones that fail quietly.
9. Parts on Condition, Not Calendar
Calendar replacement is tidy and wrong. It replaces parts that are still working and misses the ones already finished. Condition criteria are less tidy and cheaper.
| Part | Replace when | What calendar replacement misses |
| Brushes | Tufts short or collapsed; stripe left in the wash path | A brush worn flat in six weeks by grit |
| Squeegee | Lip flat, nicked, or leaving water stripes | A blade finished by a rough floor in one season |
| Filters | Airflow drops, or visible bypass at the seal | A filter loaded by one dusty campaign |
| Battery pack | All three signals agree, confirmed by a capacity test | A pack that fades early because of charge habits |
Keep the wear parts on site. A squeegee in a warehouse 900 km away turns a twenty-minute fix into a two-day outage, and the cost of that outage is not the part.
10. Four Clauses That Decide Three Years
The machine is the visible purchase; the contract is the one that keeps paying out — or not.
| Clause | What it must say | Why it matters |
| Response times by alarm level | Numbers in hours, per level, with a start point defined | "Prompt service" is not a commitment |
| Remote diagnostics and log access | Buyer can export logs and see faults without a call-out | Most faults are triaged from data, not from a visit |
| Spare and consumable list | Named parts, quantities, delivery times | Decides whether a fix is same-day or same-week |
| Data ownership | Maps, task records and logs belong to the site; full export at exit | Without it, changing supplier means an empty map |
The last one is the one sites forget. A map is months of site knowledge. If it leaves with the supplier, the next robot starts by learning a building the previous one already knew — and the saving that justified the switch disappears into the relearning.
Frequently Asked Questions
Is shorter runtime always a tired battery?
No. Check contact film and charge current first, then the depth of discharge the site actually runs. Only then does a capacity test separate a tired pack from a dirty dock.
How often should a pack be replaced?
On trend, not on age: area per shift, returns per area and time from warning to full, confirmed by a capacity test. Two months of movement in one direction justifies the test.
The robot docks at thirty per cent. Is that a fault?
Usually state-of-charge estimate drift. One calibration per the manual resets it. Do not calibrate on a routine schedule, because calibration itself costs a cycle.
Should the map be rescanned when localisation drifts?
Walk the floor first. Racks, pallets, coatings and doors change far more often than maps break. Rescan only after the physical environment is confirmed.
How often should a map be rescanned?
In a fixed weekly window, with the aisle empty and the change logged. Scanning during loading turns temporary obstacles into permanent walls.
What cleans a lidar window?
A dry lint-free cloth or lens tissue, working from the centre outward in one direction. Blow grit off first and never wipe in circles.
How often do sensors need cleaning?
Every shift in dusty sites, daily in commercial space, and after each run where oil mist is present. Log the cleaning: cleaned and provably cleaned are different claims.
Phantom obstacles keep appearing. Replace the sensor?
Clean first, then check calibration and firmware. Replacement is the last step, and rarely the answer when the fault comes and goes.
Who responds to alarms at night?
A named person with a written response target for level-one alarms. Remote viewing is not attendance: a camera cannot move the pallet or wipe the oil film.
Which contract clause matters most?
Data ownership: maps, task records and alarm logs, plus full export at exit. Without it, changing supplier means starting the site from an empty map.