Read claim 1, and note first what kind of document this is. US20260187862A1, “Feature Location Identification,” assigned to NVIDIA and published July 2, 2026, is a published application — kind code A1 — not a granted patent. The claim language below is what the applicant is seeking, not coverage an examiner has allowed, and it can narrow or change before any grant. With that framing fixed, the independent claims are worth reading for what scope they are directed to.

An autonomous or semi-autonomous machine comprising: one or more central processing units (CPUs); one or more graphics processing units (GPUs); one or more hardware accelerators; one or more image sensors; and one or more LiDAR sensors; wherein the autonomous or semi-autonomous machine is to: determine, based at least on image data obtained using the one or more images sensors, a first location associated with a road marking within an environment; determine, based at least on the first location and LiDAR data obtained using the one or more LiDAR sensors, a second location associated with the road marking within the environment; and perform one or more planning, navigation, or control operations based at least on the second location associated with the road marking.— Feature Location Identification, US20260187862A1

What claim 1 is directed to

Claim 1 is a machine claim, not a bare method. It recites an autonomous or semi-autonomous machine that includes CPUs, GPUs, hardware accelerators, image sensors, and LiDAR sensors, and it is the combination of that hardware with a specific two-step localization behavior that the claim fences. The behavior has three limitations in sequence: determine a first location of a road marking from image data; determine a second location of that marking from the first location plus LiDAR data; and perform planning, navigation, or control based on the second location. The dependency of the second location on the first is the load-bearing limitation — the LiDAR step is not an independent measurement but a refinement seeded by the camera estimate. A reading of the claim turns on that ordering: image first, LiDAR-refined second, action on the refined result.

The dependent claims that hang off claim 1 fence the refinement mechanics. Claim 3 adds determining, from the first location, a portion of the LiDAR data associated with the marking, then deriving the second location from that portion. Claim 4 adds selecting an area of the environment from the first location and using it to pick that LiDAR portion. Claims 5 and 7 introduce the LiDAR points and their intensity values as the basis for the refined location — the intensity limitation is where the retroreflective-paint physics enters the claim language. Claim 9 recites representing the first location as a top-down image and generating a second top-down image from the LiDAR data, with the refined location determined from both. Each dependent narrows how the second location is computed, without disturbing the image-then-LiDAR spine set in claim 1.

Three independent footholds, and where they land

The application carries three independent claims at different levels of abstraction. Claim 1 is the machine, quoted above. Claim 10 recites a system with CPUs, GPUs, hardware accelerators, image sensors, and depth sensors that causes a machine to act on a road-marking location determined from image data and depth data — note the generalization from “LiDAR” to “depth” sensors, which broadens the sensing modality beyond LiDAR specifically. Claim 19 drops to a system-on-a-chip: at least one SoC with CPUs, GPUs, and hardware accelerators that updates, using depth data, an initial location of an object — not just a road marking — determined from image data. Claim 19 is the broadest of the three in subject (any object, not only lane paint) while being the most specific in hardware (an SoC). Claims 18 and 20 then enumerate a long list of deployment contexts — control and perception systems, robots, data centers, LLM systems, cloud resources — which functions as claimed breadth of where the system may be embodied.

On classification, the record places the application under CPC G06T 11/00 (2D image generation) with additional G06T 2207 image-analysis sub-codes, rather than a B60W vehicle-control or G05D autonomous-navigation class. That is consistent with claims directed to determining a feature’s location within image and depth representations — a perception-and-representation placement — even though the independent claims recite the downstream planning, navigation, or control step. For anyone mapping the sensor-fusion landscape, the useful characterization is factual: this is a camera-seeds-LiDAR localization filing, claimed at three altitudes (machine, system, SoC), currently pending as a published application.

Two threads in the same July 2 drop sit adjacent to it in the perception pipeline. US20260187981A1 claims iterative classifier-based mining of training data for autonomous systems, and US20260187971A1 claims a multi-pass artifact-removal method for streamed imagery — the training and image-cleanup stages that precede a localization step like this one. A fourth, US20260187482A1, claims a language-model router — a different subject entirely, and a reminder that the same publication cycle carries claims aimed well outside the autonomy stack. None of these is granted; each is an application whose independent claims define what is being sought.

The distinction matters for how the record should be characterized. A published application marks a filing’s entry into public prosecution: its claims are on the table, examinable, and amendable, and the CPC classification is an initial placement rather than a settled one. Reporting it as a grant, or treating its enumerated deployment contexts as a description of shipped systems, would overstate what the document establishes. What claim 1 here covers is exactly the image-then-LiDAR machine recited above, no broader, and whether any given autonomy stack reads on it is a limitation-by-limitation question against a claim that has not yet issued.