Restroom cleaning robots, September 2026

Three marks run through this post. [single source] is a claim found in exactly one place and not corroborated. [secondary] is an aggregator, a blog or a data broker rather than a primary source. [inference] is our own reasoning from the stated evidence rather than a reported fact. Vendor specifications are marketing-grade unless the post says otherwise, and "not disclosed" means the vendor publishes nothing — we do not guess.

The pass ran on 7 September 2026: about 70 web searches in English, 中文 and 日本語 and about 45 full-page reads across vendor and distributor pages, Nasdaq and GlobeNewswire releases, MOM, NEA, MOH, IEC and ISO standard pages, the Restroom Association of Singapore design guide, robotics press and arXiv. Prices are list or distributor prices seen in September 2026 unless dated otherwise. The rates used are US$1 to about S$1.30 and about ¥7.1.

We looked because a competitor raised money. Hivebotics, a Singapore company spun out of the NUS Graduate Research Innovation Programme, announced a US$6 M round on 4 September 2026, led by Vertex Ventures SEA and India, to move its restroom robot Abluo from pilots into volume production and to build distributors in four markets. The question we put to the research was narrow: is a restroom cleaning robot something a 2026 full-stack team can build materially better and materially cheaper than a machine designed in 2021? The two halves resolve differently, and the answer is the reason Tidewell C1 exists and is shaped the way it is.

Headline findings

1. The incumbent machine is not a 2021 design. The unit in the field is the Abluo Mark 2, marketed as new for 2026: 3D SLAM lidar, a 6-DoF arm at 0.02 mm repeatability, six cleaning modes including 173 °C steam sanitation, a cycle of about 40 minutes, a 3-hour swappable battery, and a claimed 60 to 70% labour reduction. It buys from the same 2026 catalogue anyone else would.
2. There is no parts-cost advantage available to a second entrant. Hivebotics is a roughly 10-person company [secondary] buying the same arms, depth cameras, vacuum motors and Chinese steam OEMs. The cheap-supply-chain argument is symmetric: it applies to the incumbent on identical terms.
3. The tooling argument is partly already spent. Hivebotics did not hand-roll motion planning; it licensed MoveIt Pro from PickNik, which PickNik says saved an estimated ten months of development and testing. The 2021 to 2026 tooling gain has already been banked by the incumbent.
4. The hard-to-copy asset is the site data, not the robot. About 10,000 operating hours across some 20 sites in four regions over twelve months, and 100 or more orders [single source], is a dataset and a service history. Nothing in the parts catalogue closes that gap.
5. In Singapore the incumbent to displace is not Abluo. It is Primech AI's Hytron: NVIDIA Jetson Orin compute, a 400 mm access width, electrolysed-water jetting, door opening and lift use, greater than 99% bacterial reduction in independent testing, approved under NEA's Environmental Services Productivity Solutions Grant at up to 50% co-funding, sold on multi-year leases, with a facilities-management channel and a listed parent that has to publish. [inference]
6. The binding constraint in this product is legal, not robotic. A camera in a Singapore restroom is not a privacy preference; it is criminal-law exposure. Abluo's own description says it uses onboard cameras to identify fixtures.
7. Nobody in the category publishes operating evidence. No vendor publishes cycle time on a named site, uptime, mean time between failures, intervention rate, an ingress rating, or a third-party disinfection report behind the disinfection percentages in the marketing. Nobody maps a cleanliness score to a recognised standard, and nobody shares memory across a fleet.
8. The verdict. "Better" is available on three axes and only three. "Cheaper" is not available at all.

The incumbent, in its own numbers

Both generations are described from Hivebotics' own product pages.

Two things in that table are worth stating plainly. The widely-quoted line that thirty minutes of manual cleaning becomes a five-minute human inspection is a labour-substitution claim, not a speed claim: at 30 to 40 minutes per bathroom the robot is no faster than the cleaner it replaces. [inference] And no CE, UL or PSB mark, no IEC 63327 and no ISO 13482 file appears anywhere in the public record for this machine, nor a lab report behind its disinfection figure.

Who else is in the category

The 2026-vintage architecture in that table is Loki's, not Abluo's: an ETH Zurich spin-out doing contact cleaning with interchangeable tools, trained by teleoperation plus end-to-end learning, about US$13.4 M raised [secondary], partnered with SoftBank Robotics Japan, which sells the service and collects the training data in the region with the highest willingness to pay. Somatic anchors the category as the only vendor publishing a price. LionsBot sets what a Singapore facilities operator thinks a cleaning robot costs before the restroom conversation even starts.

The buyer's money, in five numbers

1. The wage floor is legislated and rising. Singapore's Progressive Wage Model floor for an outdoor, healthcare or restroom cleaner runs S$2,325 a month for July 2025 to June 2026, S$2,495 from July 2026, then S$2,665, then S$2,835 by July 2028-29, plus a mandatory bonus of at least two weeks. Fully loaded with CPF and that bonus, the July 2026 floor is about S$3,040 a month, roughly S$36,500 a year. [inference] That escalator is why this market is real, and it runs against a sector reported at roughly 4% net margin on a 58% labour cost base [single source, secondary].

2. The headline demand statistic is soft, and we will not repeat it. The claim that restrooms are about 6% of floor area but about 60% of cleaning hours traces to the founders and their investor, not to an independent study [single source]. The sourced substitutes are norms, not ratios: NEA allows 39 minutes 13 seconds for a 25 m² washroom; MOH's National Cleaning Standards give 45 minutes for a common toilet once a day, 30 minutes for a toilet with shower, 10 minutes for a single-room toilet. The 200 to 400% annual turnover figure quoted alongside it traces to Kimberly-Clark marketing; an ISSA and BSCAI survey found 41% of contractors reporting turnover of 50% or more and 20% reporting under 10% [secondary]. Both numbers are pitch material, and neither survives sourcing.

3. Published rental prices are low, and there are only two of them. Somatic publishes US$1,500, US$3,000 and US$4,500 a month for eight-hour weekday, sixteen-hour weekday and sixteen-hour seven-day service, which is US$8.63 an hour on weekdays. LionsBot rents floor machines at S$1,350 to 2,150 a month [single source]. Pudu's CC1 Pro floor robot is quoted from about US$479 a month [secondary]. Abluo and Hytron publish no price at all across roughly ten sources checked. Any restroom robot is quoted into a market where the buyer already has those anchors.

4. The channel and the grant are the real barrier. NEA's Environmental Services Productivity Solutions Grant co-funds up to 50% of a qualifying purchase, and Hytron is already on the approved list, which hands its buyer half the capital cost. The benchmark commercial terms in this market are Primech's: a three-year lease at a fixed monthly fee, maintenance, quarterly preventive service, corrective repair within three business days, spares, three months of onboarding free. The two large Singapore facilities channels are taken — Hytron through CBM, LionsBot through Chye Thiam.

5. Nobody has a densely used fleet. Abluo's own figures work out to about 42 hours per site per month, which is about 1.4 hours a day, or two cycles [inference]. Primech's disclosed deployments are one unit on a lease and one on a school pilot. Somatic does not disclose a unit count. No vendor in this category has published a fleet running full shifts.

Cameras: the one room where vision is the wrong answer

Closed-circuit recording in a toilet can constitute outrage of modesty or voyeurism under Singapore law, on top of Personal Data Protection Act exposure with financial penalties reaching S$1 M. That sits on top of every hospital's own procurement position. The incumbent's published description of its own product says it uses onboard cameras to identify fixtures.

The technical answer is not expensive. Occupancy is a 24 GHz mmWave presence sensor in the US$20 to 60 class, backed by a door interlock and the safety lidar's protective field. Fixture geometry and soil scoring can run on the robot and stay there, with no frame leaving the machine and an attested log saying so. The rule we adopted is that no camera-based occupancy or people-detection function exists in the machine at all, written into the safety file rather than set in software. It is the cheapest differentiation in the product and the only one with criminal law behind it, and it is a procurement gate no competitor has passed.

The standards map, which is not the one you would guess

Floor-cleaning robots are explicitly out of scope of ISO 26058-1, and they are not driverless industrial trucks, so ISO 3691-4 is the wrong hook. The right ones are IEC 63327, the first international standard for autonomous commercial floor-treatment machines — TÜV Rheinland certified Gausium against EN IEC 63327 at performance level d, and SGS certified Rosiwit — with IEC 60335-2-72, IEC 60335-2-79 Edition 5.0 for steam and high pressure, IEC 60335-2-69 for commercial wet and dry vacuum, and ISO 13482 second edition as the service-robot dossier. China's GB/T 46495-2025 for commercial cleaning robots took effect on 1 May 2026.

One surprise in the file: steam is legally cheap in Singapore. MOM requires no pressure-vessel registration where safe working pressure is at or below 0.5 bar, or where pressure times volume is at or below 100 bar-litres. A 1.5 to 2 litre boiler at 8 bar is 12 to 16 bar-litres, an order of magnitude inside the exemption. The real steam risk is scald under IEC 60335-2-79, not registration.

The risk that actually kills a first prototype is chemistry. Sodium hypochlorite rapidly attacks 304 and 316 stainless steel, aluminium is incompatible, and FKM is not recommended for it. A wetted path in PVC, CPVC, PVDF or PP with EPDM seals, with the hot steam path physically separate to the tip, is a decision that has to be frozen before any robot exists.

What "better and cheaper" resolves to

Better: yes, on three axes. A camera-free architecture, in the one room where that is a legal question rather than a preference. Published sealing, ingress ratings and a real certification dossier, in a category where nobody publishes any. And cleaning that leaves evidence behind it — a per-fixture, timestamped record that a hospital infection-control team could accept, shared across a site rather than living in one robot's app. The empty lane in this market is not a better scrubber.

Cheaper: no. The parts advantage is available to the incumbent on the same terms, the development bill for a certifiable machine is about the size of the round Hivebotics just raised, and the published rental anchors cap the rate at roughly the level the economics need to work at all. Anyone entering this category on price is entering a fight they have already lost.

That asymmetry is the whole design brief. It is why C1 is contact-capable with steam as an adjunct rather than the non-contact bet the incumbent made — non-contact was a 2021 workaround for a manipulation problem that now ships off the shelf, and scrubbing is what removes limescale, dried scale and biofilm. It is why the first site is a hospital rather than a mall, since healthcare is the only vertical where the audit record is a compliance obligation and where empanelment through CHART and RoMi-H is a gate no restroom vendor has passed. And it is why the programme is staged: it runs to month ten and stops there on one question with a numeric answer — whether the robot's cleanliness score correlates with ATP and fluorescent-marker results closely enough for an infection-control team to accept it in place of a manual visual audit. If it does not, we stop.

What we will not claim

No disinfection percentage without a named lab, an organism panel and a report. No cycle time until it is measured on a real restroom, with the date. No labour-reduction percentage — the 60 to 70% figure in this market is a vendor's claim about its own machine. No price until an offer exists. No named customer or site without written permission. And no camera-free claim in our marketing until the architecture has been built and audited by a third party; until then it is design intent, and the product page says so.

Open questions

Six things are unresolved and sit ahead of any spend: a freedom-to-operate patent search against Hivebotics, Primech, Somatic and Loki; the actual price points behind Abluo's and Hytron's silence; the ingress class Singapore and EU buyers will require; whether operating robots as a service needs our own NEA cleaning business licence; RoMi-H empanelment criteria and timeline for a cleaning robot, which no vendor has been through; and whether the method in SS 694:2023 can be machine-scored at all, and whether NEA or the Restroom Association would accept robot-generated evidence against it.

Where the programme stands is on the roadmap, and what the machine is meant to be is on the C1 page. Both change when a measurement exists, not before.