The limitation that does the work in Apple Inc.'s newly published application is not a rendering trick, a pairing protocol, or a camera geometry. It is a binary: alive, or not alive. US20260202514A1, titled "MACHINE LEARNING BASED OBJECT IDENTIFICATION" and naming inventors Ke-Yu Chen, James T. Curran, Jun Gong and Gierad Laput, published on July 16, 2026. Every independent claim in it runs through the same four operations and terminates on the same step — deciding whether a detected object is a living object or a non-living object. Before anything else is said about it: this is a published application. Nothing in it has been granted, and the scope described below is the scope applied for, not the scope allowed.

Claim 2 is the first operative claim in the published text; claim 1 appears as "(canceled)." Claim 2 is directed to a system performing "object classification at a first electronic device," reciting a sensor, one or more processors, and one or more memories holding instructions. The operations proceed in order. First, the first device generates, using its own sensor, sensor data covering "a plurality of objects in a field of view of the sensor." Second, it obtains from a second electronic device that device's sensor data covering the same plurality of objects. Third, it transforms both datasets into a single body of transformed data. Only then does the fourth and final step run, and that final step is the one that defines the invention:

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

Two further independent claims carry the identical endpoint in different statutory dress. Claim 9 is a method claim, reciting the same generate-obtain-transform-classify sequence performed by a first electronic device. Claim 16 is directed to "a non-transitory computer-readable medium storing computer-executable instructions" that cause an electronic device to perform the same four operations. System, method, article of manufacture — three categories, one endpoint. Where an applicant writes the same terminal limitation into all three independent claims, that limitation is the thing being claimed, and here it is the living-versus-non-living determination rather than the multi-device fusion that precedes it.

The abstract describes one thing; the claims claim another

This record is a clean illustration of why the claim set and the abstract are not interchangeable. The abstract describes the disclosure in fusion terms. It opens by explaining that "Embodiments described herein provide techniques to enable spatial sensor data from multiple devices to be fused into a single coordinate space," and goes on to describe processing that fused data into a point cloud, transforming the point cloud to enable classification, and extracting features within multiple feature spaces. Its closing sentence describes using object classification to correlate objects detected by "multiple sensor equipped devices" — the phrase is printed that way, without the hyphen — in order to determine a coordinate space transformation between those devices.

Read that abstract alone and you would come away describing a cross-device registration invention: two gadgets reconciling their frames of reference. The claims do not support that reading. The living-versus-non-living determination that closes claims 2, 9 and 16 is absent from the abstract entirely, and coordinate-space transformation — the note the abstract ends on — never appears in an independent claim at all. The abstract is description of what the specification discloses. The claims are the measure of what is sought. On this record they point in visibly different directions, and only one of them is enforceable subject matter.

The dependent claims fill in the machinery of that classification step rather than redirecting it. On the method side, claim 10 recites clustering, "using a trained clustering model, points in the transformed data to determine an area of interest" containing the object. Claim 11 recites "analyzing, using a trained classifier model, a set of features extracted from the transformed data to identify the at least one object" — the trained-model recitation that supplies the "machine learning based" of the title. Claim 12 adds a pre-filtering step, narrowing the plurality of objects out of "a first set of multiple objects in the first sensor data" before fusion. Claim 15 specifies the data structures, calling for "a two-dimensional representation of first point cloud data in the first sensor data" alongside an image-space transformation of a graphic of the second device's point cloud. Claims 17, 18 and 19 mirror that same progression on the computer-readable-medium side.

Coordinate-space registration does appear — as a dependent option, well downstream. Dependent claim 7 recites "determining a coordinate space transformation between a first coordinate space associated with the first sensor data and a second coordinate space associated with the second sensor data," and claims 14 and 21 carry the same recitation into the method and medium families. That is a narrow add-on to a relative-position determination, not the claimed invention. One further procedural note worth recording precisely: because claim 1 is canceled, any characterization of this application that begins "claim 1 requires" has no text behind it. The first operative claim is claim 2.

Where the filing lands in the classification map

The CPC assignments are the most informative single fact about this record's landscape position, and they are not where a reader would guess. The application is classified under G01S 7/412, G01S 13/42 and G01S 13/89 — three radar groups, covering the processing of received radar echo signals, radar direction-finding, and radar imaging and mapping respectively. It is not classified into the computer-vision or neural-network groups that the title's "machine learning" phrasing might suggest. The claims themselves recite only "a sensor" in generic terms, without naming a modality, but the classification places the disclosure squarely in radar signal processing art. That combination — generic sensor language in the claims, radar classification on the cover — is worth noting for anyone mapping this filing against the surrounding art, because prior-art searching will run through the radar groups regardless of the claim's silence on modality.

The application sits in a period of dense publication activity for the assignee. Apple Inc. records published alongside it include US20260189472A1 on artificial-intelligence model coordination between a network and user equipment, US20260195972A1 on image rendering using triangle primitives, US20260194736A1 on a camera including two light-folding elements, US20260186371A1 on an electrical flexure component for sensor-shift cameras, and US20260189875A1 on location data harvesting and pruning for wireless accessory devices. Those are optical and wireless-positioning filings; the present application approaches scene understanding from a different classification neighborhood entirely.

For a docket-tracking purpose, the posture is straightforward. US20260202514A1 is an A1 publication dated July 16, 2026. Prosecution has not concluded, the independent claims as printed may be amended before any allowance, and no exclusionary right exists on this disclosure today. What the record does establish, and establishes clearly, is what Apple Inc. has asked for: a system, a method and a medium in which two devices pool what their sensors see of the same group of objects, and the pooled data is used to answer one question about at least one of them — living, or not.