NVIDIA (NVDA) is best known to investors as the company selling the compute for everyone else's autonomy ambitions. The patent record issued the week of June 9-16, 2026 shows it is also assembling coverage further up the stack - in the perception software and training tooling that turn raw sensor data into a drivable scene. In one search of grants tagged to autonomous-vehicle, robot, LiDAR and navigation themes, 33 patents issued across the sector that week; NVIDIA accounted for roughly nine of them, the single largest share, ahead of LG Electronics at three and a long tail of one-offs from Ford, Hyundai, GM, Bosch, Waymo and others.

The most directly autonomy-relevant grant in the batch is US12651465B2, "Multi-view deep neural network for LiDAR perception." Its claims describe a chained neural-network pipeline that processes the same 3D scene from more than one viewpoint - a perspective view and a top-down view - to produce bounding boxes and class labels for detected objects. The CPC classification is unusually dense for a software patent, spanning G05D 1/0088 and G05D 1/81 (control of autonomous vehicles), B60W 60/0011 and related conjoint-control classes, and G01S 17/89 and G01S 17/931 (LiDAR for vehicle ranging). That combination places the claim squarely at the junction of sensing and vehicle control rather than in generic computer vision.

As such, the techniques described herein may be used to detect and classify animate objects and/or parts of an environment, and these detections and classifications may be provided to an autonomous vehicle drive stack to enable safe planning and control of the autonomous vehicle.- Multi-view deep neural network for LiDAR perception, US12651465B2

What the surrounding grants cover

The LiDAR patent does not sit alone. US12657819B2, "Surface estimation," issued June 16, covers training a deep neural network to predict a dense 3D representation of a road surface, including synthetic generation of 3D road surfaces with varied direction and lateral slope. Its CPC tags - B60W 30/09, B60W 40/06, B60W 60/001, plus G06V 20/05 and G06V 20/58 - point to road-surface understanding feeding directly into vehicle planning. Read together with the LiDAR detection grant, the two describe complementary halves of a perception problem: what is on the road, and the shape of the road itself.

A second sub-cluster concerns the camera image-signal chain that perception models depend on. US12657818B2, "Distortion correction for environment visualizations with wide angle views," covers a Panini-projection method for correcting distortion in wide-angle vehicle camera feeds. US12651424B2, "Image harmonization for image stitching systems and applications," covers blending rendering parameters across the seams where multiple camera images are stitched into a single surround view. Both are the kind of unglamorous pre-processing that determines whether a downstream detector sees a clean scene or an artifact-laden one, and both are now issued claims.

The third strand is training data itself. US12651480B2, "Data set generation and augmentation for machine learning models," covers using a generative neural network to synthesize samples for the specific attributes where a model is underperforming, then folding them back into the training set. This is a claim over the feedback loop of model improvement rather than over any single model - coverage on the tooling, not just the output. It connects to the synthetic road surfaces in the surface-estimation grant and to NVIDIA's broader generative-3D work, exemplified the same week by US12657822B2, "Generative modeling for 3D objects using neural implicit representations."

Reading the footprint

Taken as a set, the week's NVIDIA grants map onto a recognizable shape: perception inputs (LiDAR multi-view detection, road-surface estimation), the camera-image conditioning that precedes them (distortion correction, stitching harmonization), and the synthetic-data machinery that trains the lot. The CPC overlap is notable - B60W 2420/403, which tags LiDAR as a control input, recurs across several of these grants, and the perception filings repeatedly recite the "control system for an autonomous or semi-autonomous machine" framing in their claims.

For a company whose disclosed automotive position has historically been told through its Drive platform and chip roadmap, the issued claims indicate where NVIDIA is also building defensible coverage in software. The filings point to investment not only in the perception models a customer would run, but in the data-generation and image-conditioning steps around them. Competitors filing in the same week - Waymo's US12654717B2 on trajectory-aware gear shifts, Bosch's US12657357B2 on 6D object-pose estimation, Ford's tagged-object tracking grant - each landed a single issued patent in this search; NVIDIA's cluster of roughly nine spans more of the pipeline end to end. The record describes a company extending its presence from the silicon into the perception and training layers that sit on top of it.