A granted patent is enforceable coverage, not a press release, so the useful question about a consumer-robotics company is where its issued claims actually sit. Roborock makes floor-cleaning robots, and in the U.S. grant cycle dated May 12, 2026 it issued four patents that look exactly like what you would expect from a vacuum maker. But the four-grant block is only the surface. Underneath it is a footprint of more than fifty issued U.S. patents whose center of gravity is not the brush — it is the navigation stack that lets the machine find its way around a room without a person steering it.

Start with the week's grants, because they set the frame. US12622563B2 covers an automatic cleaning device combining a dry-cleaning module and a wet-cleaning module with a detachable driving-and-supporting platform. US12622558B2 covers a detachable brush with a "foolproof structure for preventing missing mounting," and US12622559B2 covers a cleaning device and its control method. These are mechanism and serviceability claims — the physical hardware a buyer sees. The fourth grant already starts to move up the stack: US12622560B2 covers a "breakpoint continuous mopping" method, in which a robot that runs out of water mid-task remembers where it stopped and returns to finish the unmopped region after refilling.

the mopping robot is switched to a mopping mode after the water tank is supplemented with water, and carries out supplementary mopping on an unmopped region including the first breakpoint position.— Breakpoint continuous mopping method and apparatus for mopping robot, medium and electronic device, US12622560B2

That breakpoint-resume claim is a tell. Remembering an unfinished region and routing back to it is not a cleaning-hardware problem; it is a state-and-mapping problem. And mapping is exactly where the rest of Roborock's issued footprint lives.

The footprint reaches into the autonomy stack

Across the company's broader grant record — more than fifty issued U.S. patents — the classification mix moves decisively from the A47L cleaning classes into the G05D and G06T classes that cover autonomous navigation and machine vision. US12620130B2 covers a locating method for a robot: determining "current possible pose information" from ranging data, then matching current camera images against historical images and historical poses to resolve which pose is correct — visual relocalization, classified in G05D 1/0246 and G06V 10/70. US12608006B2 covers a stereo ranging method that computes object depth from the parallax between two image-collection devices, the kind of depth perception that lets a robot judge distance to an obstacle.

The obstacle and mapping coverage runs deeper still. US12572148B2 covers detecting an obstacle by transforming sensor data into depth information, converting it into a point-cloud map, and determining whether an obstacle falls within a valid analysis range in the height direction — point-cloud obstacle detection, classified across G06T 7/11, 7/50 and 7/85. US12650694B2 covers a map-drawing method that scans a region boundary, merges coordinates and divides the result into sub-regions to draw a detailed map (G05D 1/2246, G01C 21/383). And US12541936B2 covers a near-field object-detection method that shoots two images before and after a fill light fires and compares them to find a close obstacle "without adding additional apparatuses" — a claim that gets perception out of existing hardware.

The dates on the footprint matter as much as the classes. Roborock's issued grants span several years — the broader record shows filings reaching grant across 2022 through 2026 — with the navigation-and-vision patents concentrated in the most recent cohorts. The locating, ranging and near-field-detection grants all carry 2026 issue dates, and the map-drawing grant is among the freshest of all. A portfolio that keeps issuing fresh claims in localization and perception, cycle after cycle, is one whose coverage in those areas is still being built out rather than coasting on older filings — the pattern of a company actively extending its fence around the autonomy layer, not merely maintaining it. The cleaning-mechanism grants, by contrast, read as steady incremental coverage on a mature product surface, the routine maintenance of an established hardware line rather than the opening of a new front.

Where the coverage matters

The two layers tell different parts of the same story. The week's four grants fence the consumer-facing hardware — the cleaning modules, the serviceable brush, the resume logic — which is the part of the product a buyer evaluates in a store. The broader footprint fences the autonomy: localization, stereo depth, point-cloud obstacle detection, region mapping and low-cost near-field sensing. That second layer is the harder thing to replicate, because it is the software-and-perception core that determines whether a robot actually cleans a real home without getting stuck, and it is where a new entrant has the furthest to travel. A grant in A47L covers a way to build a mopping head; a grant in G05D and G06T covers a way to make the machine know where it is and what is in front of it.

None of this maps to a market outcome, and an issued patent describes coverage, not a shipped feature — some of these methods may sit unused. What the record shows is concrete: Roborock's enforceable footprint spans both the mechanism and the navigation, and the navigation half — pose estimation, depth, point clouds, mapping — is the larger and more technically loaded portion. For a reader sizing up where a consumer-robotics company has actually locked in coverage, the brush is this week's headline, but the autonomy stack is the body of the portfolio, and it is already issued rather than pending.