A granted patent is enforceable coverage, so a cluster of grants issuing in one week is a snapshot of where a company has just secured territory. In the week of April 28, 2026, Nvidia Corporation had a run of patents issue, and a recognizable group of them lands not on chips or cooling but on something more foundational to robotics and autonomy: how a machine is trained to see. For a company that positions itself as the compute-and-software supplier to the entire field rather than a vehicle or robot maker, issued coverage over perception training is coverage one layer beneath everyone building on its stack.

The clearest pair is US12614382B1 and US12614380B1, two grants both titled neural network based vision systems and both claiming ways to train vision networks on unannotated images using loss functions computed from image patches.

Apparatuses, systems, and techniques to train one or more neural networks using unannotated images.— Neural network based vision systems, US12614382B1

Training on unannotated images matters commercially because labeling is the expensive bottleneck in building any perception system — the cost that scales with every camera-equipped robot or vehicle. Two same-day grants over methods that reduce the need for hand-labeled data give Nvidia issued coverage at exactly that chokepoint, classified in G06V 10/82 (neural-network image recognition) and G06N 3 machine-learning classes.

What the grants fence off

The labeling theme recurs from the other direction in US12614284B2, on class-agnostic object mask generation. It claims an auto-labeling framework that takes images and bounding boxes and produces pixel-level segmentation masks usable to train further models — coverage over the manufacture of training data itself. US12614062B2, on classification with neural networks, claims training classifiers using compressed representations of class labels with fewer bits than the labels themselves — an efficiency method aimed at the same training pipeline. Across these grants the common subject is the upstream work of preparing data and training the networks that later perceive a scene.

The application end of the spectrum shows up in US12614457B1, an AI-optimized parking system that uses computer vision and a large language model to analyze a vehicle and its occupants, then recommend and route to a parking space. Classified in G08G 1/141 and G08G 1/146 (parking-space guidance) alongside G06V vision classes, it is the one grant in the group that reaches all the way to a driving-adjacent use case — a reminder that the same perception toolkit Nvidia is patenting upstream also surfaces in concrete vehicle applications.

The shape of the cluster is what gives it commercial weight. Three of the grants — the two unannotated-image training patents and the auto-labeling patent — address the data side of building a perception system: how to train when labels are scarce, and how to generate labels when you need them. The classification grant addresses the training side: how to do it more efficiently. The parking grant shows the output side: a working system that perceives a scene and acts on it. A footprint that spans data preparation, training method, and application is broader than any single claim, because it touches several distinct steps a downstream builder must pass through. For Nvidia specifically, whose stated role across robotics and autonomy is to supply the training-and-inference platform rather than the finished robot or car, coverage that sits at the data-and-training layer reaches a wide set of customers regardless of what those customers ultimately ship. The same toolkit underlies an autonomous truck's perception, a warehouse robot's bin-picking vision, and a humanoid's scene understanding; none of those products needs to be Nvidia's for the underlying method to be Nvidia's.

The footprint and the field around it

What this buys, in business terms, is coverage over technique rather than product. A patent on a way to train a vision model without labels, or to auto-generate segmentation masks, reads on a process that sits underneath autonomous-vehicle perception, warehouse-robot vision, and humanoid sensing alike — regardless of which company ships the end device. That is consistent with how Nvidia is positioned across the sector: as the supplier of the training and inference layer rather than a competitor in any single robotics product market. The week's broader grant record underscores the surrounding density — in the same issue week the autonomous-vehicle grant set was led by volume from Nvidia and Toyota, with Hyundai, GM, Zoox and Ford all among the named assignees, and machine-learning and vision classes (G06N, G06V) recurring throughout. The whole field is filing in this taxonomy; Nvidia's grants sit at the training-infrastructure end of it.

Read as a footprint, the cluster is coherent: multiple same-week grants over how vision networks are trained and how their training data is produced, plus one that carries the toolkit into a parking application. None of this is a sensor patent or a vehicle patent; it is coverage over the methods that turn pixels into machine perception. For a company whose robotics and autonomy revenue comes from selling that capability to others, issued coverage upstream of the application is coverage with broad reach across the customers building downstream.

The limits are worth stating plainly. A grant records what a patent office allowed, not how widely a claim will be practiced or whether it will ever be asserted, and many issued patents are never litigated. Several of these grants are also general machine-learning methods whose relevance extends well beyond robotics. What the week's record shows is specific and bounded: in a single issue week, Nvidia added patents whose claims cover training vision neural networks, auto-labeling image data, efficient classification and a vision-driven parking system — mapping its freshly issued footprint onto the perception-training layer that the robotics and autonomy field builds on top of, in a classification space crowded with the rest of the sector.