Every flagship robot vacuum launched in the past eighteen months has been sold on artificial intelligence. Roborock says its Saros Z70 recognizes 108 pre-programmed objects and lets you label 50 more; newer Roborock marketing claims "up to 301 types of objects recognizable." Ecovacs sells the Deebot X11 with Yiko, described on its own store as "a fully autonomous AI agent butler that uses deep thinking." Narwal's CES 2026 pitch centered on NarMind 2.0, "a vision-language model, Omni Vision AI, and dual RGB cameras." Dreame claims 240 to 280-plus recognizable objects.
Almost none of it is tested. That is the story.
The claims nobody checks
There is a structural gap in this market. Vendors compete on AI features. The independent test houses — the ones with repeatable methodology — measure suction, mopping and obstacle avoidance. They do not measure the AI features at all.
Vacuum Wars bought and reviewed the Ecovacs X11 OmniCyclone and ran no dedicated test of Yiko or its on-device AI. Gizmodo covered the launch and explicitly declined to validate the agentic claims without hands-on data. TechRadar, reviewing the same category, noted that "the term 'agentic' is being used quite loosely, and we're not actually looking at a robot vacuum that's going to make truly independent decisions."
Across every vendor product page we read, not one published a TOPS rating or named a silicon supplier for the "on-device AI" doing this work. The single attribution of a dedicated NPU in the X11 traces to a retailer blog, not to Ecovacs.
Our view: an unfalsifiable claim in a category where nobody tests is not a specification. It is a marketing asset, and it should be priced as one.
Where the claim collided with a tester
The exception — and it is instructive — is Roborock's OmniGrip arm on the Saros Z70, because a physical arm picking up a physical sock is something a reviewer cannot avoid evaluating. Three independent testers reached the same verdict.
Vacuum Wars found the arm "hit or miss… sometimes it would pick up the objects, and sometimes it wouldn't," working roughly half the time and recognizing only clumps, fabrics and shoes — "limited to the items it was specifically trained on by Roborock." TrustedReviews watched a sock and a tissue get "completely ignored," not even registering as objects on the map, and reported the claw getting caught inside a shoe and lifting it back off; its conclusion was that "most people would be better off buying a regular robot vacuum and mop." MamanVsRobot described it as slow and "limited to a few categories."
Against a vendor claim of 108 object types, three testers measured roughly coin-flip reliability on three or four.
Here is the part that matters for buyers. The same reviewers rate the underlying vacuum highly — the Z70 scored 22 out of 24 on Vacuum Wars' obstacle-avoidance test against a tested average near 16.6, and TrustedReviews confirmed cables, fake pet mess and shoes were "easily avoided." The AI navigation works. The AI manipulation does not. Those are two different products sharing a marketing word.
Object counts do not predict performance
The clearest evidence that the numbers race is decorative is that the numbers are anti-correlated with the results.
On Vacuum Wars' 24-point obstacle course, Eufy's Omni S2 (claiming ~200 objects) and Roborock's Saros 10R (Roborock claims up to 301) both scored a perfect 24/24. Ecovacs' T90 Pro Omni and X12 OmniCyclone scored 23. The X11, from the brand selling the "AI agent butler," scored 20. And Gizmodo found Narwal's camera-only system — claiming just 120-plus objects, with no time-of-flight sensor at all — beat the ToF-equipped Roborock on avoidance.
Vacuum Wars named the dynamic directly: the industry is in "a race over the number of recognizable objects." Nothing in the measured data suggests the winner of that race cleans better.
What does correlate, per the same test house, is whether a camera exists at all: "robots that rely solely on structured light without a camera or IR laser system have been among the lowest-scoring models we've evaluated." And the one concrete failure mode a tester actually isolated is invisible to any object count — Narwal "did still suck up a couple of charging cords… on my patterned rugs. I suspect the camera needs more contrast."
The premium is collapsing on its own
The market is already correcting. The Saros Z70 launched at $2,599.99 MSRP against the armless Saros 10R at $1,599.99 — a $1,000 premium for the arm. Android Authority's assessment was blunt: "does the OmniGrip arm help the Saros Z70 actually do what it's supposed to do — clean your floors — to the tune of a $1,000 price hike? The short answer is no."
By July 2026 the Z70 sells for $999.99, a 62% cut from MSRP, at or below the price of the armless model it was supposed to sit above. The Ecovacs X11 fell from a $1,099 observation in March 2026 to $699.00 in July. Vacuum Wars' own July ranking is led not by any flagship but by the $899.99 Dreame L60 Ultra PE, with a $563 Ecovacs T90 Pro Omni outranking a $1,699 Dreame. Their measured summary: premium models score just "2–5% higher on carpet deep-clean than the best mid-range options."
Read the shape carefully, because it is easy to overclaim. The flagship *price band* is stable year over year — Dreame's X60 launched straight into the $1,499–$1,699 slot the X50 vacated, Ecovacs holds the X12 at $1,499. What collapses is each individual model the moment it is superseded. That is fast product cycles and inventory clearing at least as much as it is buyer rejection of AI pricing.
The regulatory and privacy overhang
The FTC has been pursuing exaggerated AI capability claims since Operation AI Comply launched in September 2024, and enforcement continued into 2026 with an action against CMG Media over a fictitious tool. No action against a robot vacuum maker has been brought — the relevance here is by analogy, not precedent.
The privacy record is less hypothetical. MIT Technology Review documented gig workers leaking at least fifteen images from iRobot development units in 2022, including a woman on a toilet. At DEF CON 32 in 2024, researchers demonstrated Bluetooth takeover of Ecovacs units from roughly 450 feet, with camera and microphone access and no hardware recording indicator — and showed that deleting your account removes neither data nor auth tokens. Eufy was caught in 2022 uploading footage to cloud servers without encryption despite local-storage marketing.
That last one is the direct precedent for today's unverifiable "on-device AI" claims. The same company, making the same class of unauditable processing claim, was already wrong once.
What a buyer should actually do
Pay for the camera, not the object count. Treat the AI navigation as mature and the AI arm and AI conversation as demos. And treat "on-device AI" as an open question until a vendor names the chip — because on current evidence, nothing in this market checks whether it is true.
Sources
- Roborock Saros Z70 product page — vendor claims: OmniGrip arm, 108 objects, 50 user-defined
- Roborock StarSight — vendor claim on sampling frequency
- Ecovacs Deebot X11 OmniCyclone — "AI agent butler" claim
- Vacuum Wars: Roborock Saros Z70 review — arm works about half the time
- TrustedReviews: Roborock Saros Z70 — sock and tissue ignored; shoe caught in claw
- MamanVsRobot: Saros Z70 review
- Vacuum Wars: best robot vacuums for obstacle avoidance — 24-point scores
- Gizmodo: Narwal Freo Z Ultra review — camera-only beats ToF; cords on patterned rugs
- TechRadar on "agentic" robot vacuums
- Android Authority: Saros Z70 review — the $1,000 premium verdict
- Vacuum Wars: best robot vacuums, July 2026 — 2–5% carpet gap; current ranking
- Vacuum Wars: CES 2026 coverage — Narwal NarMind 2.0, Dreame and Roborock arm/leg claims
- Benesch: one year of FTC Operation AI Comply
- MIT Technology Review: Roomba images leaked
- TechCrunch: Ecovacs robots hacked at DEF CON 32
