Apple Inc. is applying for coverage on a sensing system whose stated endpoint is not mapping a room or recognizing a face, but deciding whether the thing a device just detected is alive. Published on July 16, 2026 as US20260202514A1, MACHINE LEARNING BASED OBJECT IDENTIFICATION names inventors Ke-Yu Chen, James T. Curran, Jun Gong and Gierad Laput, and is classified under G01S 7/412, G01S 13/42 and G01S 13/89 — radar-side classification codes rather than the image-processing buckets where most consumer-device vision work lands. For a business desk, the classification is the first tell: this is a filing about what a device can infer from radio returns, not from a camera feed.

The operative scope sits in claim 2. Claim 1 is marked (canceled) on the published record, so claim 2 is the first operative independent claim, and it recites a system at a first electronic device — a sensor, one or more processors, memories — performing four steps in sequence. It generates its own sensor data on a plurality of objects in the sensor's field of view. It obtains second sensor data covering the same objects from a second electronic device. It transforms both datasets into common transformed data. And then it classifies. Claim 9, a method claim, and claim 16, a non-transitory computer-readable medium claim, recite the same four steps in different statutory dress and land on the identical final limitation.

That final limitation is the invention, and it reads the same way in all three independent claims:

classifying at least one object of the plurality of objects based on the transformed data for the plurality of objects to determine whether the at least one object is a living object or a non-living object.— MACHINE LEARNING BASED OBJECT IDENTIFICATION, US20260202514A1

The commercial reading of that sentence is narrower and more specific than a generic sensor-fusion story. A living/non-living decision is the primitive that ambient computing has been missing at the low-power tier. Cameras answer it well and carry privacy, power and placement costs that keep them out of most rooms. Radar-class sensing answers presence cheaply but has historically struggled to distinguish a person from a chair, a pet from a fan, or a moving occupant from a moving curtain. An application directed to producing that distinction from fused, multi-device sensor data is an application about making presence detection trustworthy enough to act on — and acting on presence is what determines whether a home device wakes, hands off, dims, alerts or stays silent.

The second device is the load-bearing element

What separates this filing from a single-sensor classifier is that the second device is not optional color — it is recited in every independent claim. The system does not merely use its own sensor; it obtains second sensor data from a second electronic device covering the same plurality of objects, and the classification runs on the fusion of the two. Strategically, that is a filing that presumes a household with more than one Apple-designed sensing endpoint in the same room, and treats the presence of the second one as an input rather than an accessory. The installed base becomes the sensing array.

The dependent claims fill in the machinery in a way that reads like an implementation roadmap rather than a defensive net. Claims 3, 10 and 17 add a trained clustering model that groups points in the transformed data into an area of interest. Claims 4, 11 and 18 add a trained classifier model analyzing features extracted from that transformed data. Claims 5, 12 and 19 add pre-filtering, narrowing a larger set of detected objects down to the plurality that matters. Claims 8 and 15 specify two-dimensional representations of point-cloud data. Claims 6, 13 and 20 determine the relative position of the second device to the first by comparing the paths of a commonly detected object across both datasets — inferring where the devices sit relative to each other from what they both happen to be watching, with no setup step for the user.

Only after that, at dependent claims 7, 14 and 21, does the filing reach coordinate-space transformation between the two devices' sensor frames. It is worth being precise about the hierarchy, because it inverts the intuitive reading: the cross-device geometry is a downstream dependent option, and the living/non-living decision is the terminal limitation of every independent claim. A reader working from the abstract alone would arrive at the opposite conclusion. The abstract describes the disclosure as techniques enabling spatial sensor data from multiple devices to be fused into a single coordinate space, processed into a point cloud, and used to correlate objects across "multiple sensor equipped devices" — the published text carries that missing hyphen — and closes on coordinate-space transformation. The living/non-living classification that ends claims 2, 9 and 16 appears nowhere in that description.

Where it sits in the July 16 cohort

The application does not stand alone in Apple's July 16 publications, and the surrounding set suggests a company building the plumbing that a multi-device sensing layer would require. ARTIFICIAL INTELLIGENCE MODEL COORDINATION BETWEEN NETWORK AND USER EQUIPMENT is directed to coordinating AI models between the network and user equipment. Location Data Harvesting and Pruning for Wireless Accessory Devices concerns how location data is gathered and pruned across accessory devices — the same many-devices, one-picture problem, approached from the location side. On the capture side, Camera Including Two Light Folding Elements and Electrical Flexure Component, Separate From Suspension, For Sensor Shift Cameras continue the optics and stabilization work, while TECHNIQUES FOR IMAGE RENDERING USING TRIANGLE PRIMITIVES and TECHNOLOGIES FOR RADIO LINK CONTROL RETRANSMISSION BASED ON PACKET IMPORTANCE cover rendering and radio-link retransmission respectively. Read together, the cohort is less a single product signal than evidence of sustained investment across sensing, on-device inference and the links between devices.

Two drafting details are worth flagging for anyone reading the record directly, because they affect how the document should be cited. Five dependents — claims 3, 4, 5, 6 and 8 — still recite "The system of claim 1" even though claim 1 is marked canceled, leaving those five formally dependent on text that no longer exists. And claims 13 and 20, a method claim and a medium claim respectively, both recite one or more memories storing instructions executed by one or more processors — but the parent claims do not supply both terms. Claim 9 recites neither memories nor processors, so claim 13's reference to each of them is unsupported. Claim 16 does recite one or more processors, so claim 20's processors are accounted for; it is the memories, absent from claim 16, that claim 20 refers back to without antecedent basis. These are the kinds of artifacts that get resolved in prosecution; they are noted here only because they mean the phrase "claim 1 requires" has no support on this record.

The final caveat is the most important one for anyone treating this as a signal. US20260202514A1 is an A1 publication — a pending application, not a granted patent. Publication confirms that Apple filed this scope and that the disclosure is now public; it confirms nothing about what will issue, and the claims as published may be amended or narrowed before any grant. What it does establish, on the record, is where the engineering effort went: not into a better point cloud for its own sake, but into a system that pools what two devices see and returns an answer about whether something in the room is alive.