Is the robot fleet actually funded? For NVIDIA the better question is whether the autonomy stack on top of its chips is actually owned — and a June 9, 2026 grant says it increasingly is. US12651465B2, "Multi-view deep neural network for LiDAR perception," claims a neural network that fuses multiple views of LiDAR point-cloud data into the object detection and scene understanding a self-driving system needs.

Read the segment, not the keynote. NVIDIA's automotive and robotics revenue is small next to its data-center business — the company's own FY2025 10-K reported Automotive revenue of about $1.69 billion against $130.5 billion in total revenue, roughly 1.3% of the company. But its strategic value is that NVIDIA does not just sell the inference chip; it sells the perception software that runs on it. This patent's CPC fingerprint makes that explicit: G05D 1/0088 (autonomous vehicle control), B60W 60/0011 (autonomous driving maneuvers), G01S 17/931 (LiDAR for driving), and a stack of G06V computer-vision tags. That is a full perception pipeline, not a component.

“A deep neural network(s) (DNN) may be used to detect objects from sensor data of a three dimensional (3D) environment. For example, a multi-view perception DNN may include multiple constituent DNNs or stages chained together that sequentially process different views of the 3D environment.”— U.S. Patent No. 12,651,465 source

The claims are where the breadth shows. The grant runs to 20 claims across three independent forms — "one or more processors" (claim 1), a "method" (claim 9), and a "system" (claim 17) — the standard belt-and-suspenders structure that lets a holder assert the same invention against a chipmaker, a software vendor, and a full-system integrator alike. The technical core is a two-stage pipeline: claim 1 recites receiving "initial classification data generated using a first representation of sensor data," then generating "refined classification data" by running a neural network over "a projected representation of the initial classification data and a second representation of the sensor data different from the first." In plain terms, the patent covers processing one view of the LiDAR point cloud, projecting those labels into a second view, and refining them — the multi-view trick named in the title.

The dependent claims map the moat's edges. Claim 7 specifies the two representations as "one or more range images corresponding to a point cloud" and "one or more height maps corresponding to the point cloud" — the perspective-view and top-down-view pairing that makes the method concrete. And claim 8 is the deployment net: it recites that the processors may sit in "a control system for an autonomous or semi-autonomous machine," "a perception system," a system "implemented using a robot," one "implemented at least partially in a data center," or "using cloud computing resources." That single dependent claim deliberately spans the car, the robot, the edge box, and the cloud — every surface on which NVIDIA already sells compute.

The middle claims show the holder has fenced the implementation details, not just the high-level idea — the kind of specificity that makes a claim hard to design around. Claim 5 recites that the network processes "one or more first channels comprising one or more height values" alongside "one or more second channels comprising one or more view transformed classifications," and claim 6 reaches "one or more view transformed confidence maps or one or more view transformed segmentation masks." In other words, the patent does not merely cover "fuse two views" in the abstract; it covers the specific representations — height channels, projected confidence maps, segmentation masks — an engineer would actually reach for to build a working multi-view LiDAR detector. That is the difference between a claim a competitor can route around and a claim that sits on the natural path to the result.

The arms-dealer thesis is a cash-flow thesis. A company that owns the chip and the perception model captures margin at two layers and locks customers into a platform. Every autonomy startup that builds on NVIDIA's perception IP is a customer that is hard to migrate away, and a claim set written to cover "an autonomous or semi-autonomous machine" or "a robot" running the method is a claim set written to follow that customer wherever it deploys. The patent is one more switching cost being manufactured and put on the public record.

For a fundamentals-driven reader, the move to track is whether NVIDIA's automotive/robotics segment disclosure starts reflecting software-attach economics rather than unit chip sales. Patents like this one are the leading indicator — the IP shows up years before the revenue mix shifts, and right now the entire Automotive line is buried inside the Compute & Networking segment alongside the data-center business. R&D capitalized into a perception-software moat is R&D that the income statement won't fully reveal for a while; the patent record is where it surfaces first.

The honest limit: a perception patent is a method, not a market share. A 20-claim grant on multi-view LiDAR perception does not tell you how many autonomy programs license NVIDIA's stack or at what price, and a granted claim is not the same as an enforced one. It tells you NVIDIA is deliberately accumulating IP at the software layer of autonomy, written broadly enough to reach cars, robots, edge devices and the cloud, consistent with a strategy of being indispensable to everyone else's machine.

The takeaway for the autonomy money desk is structural. When the dominant compute supplier also owns the perception software — and writes the claims to cover every form factor that software might run on — the question for every other autonomy name becomes "how much of your stack is really yours?" This grant is NVIDIA quietly tightening that question into a contract term.