Nvidia doesn't want to win the robotaxi race. It wants to sell the perception stack to everyone running in it. The grant US12651465B2, "Multi-view deep neural network for LiDAR perception" (issued June 9, 2026), is what that strategy looks like as a patent claim.

The mechanism, plainly: raw LiDAR returns are a sparse 3D point cloud, awkward for a neural network to read directly. The patent's approach is to project that cloud into multiple views — a top-down bird's-eye view, a perspective view — and run a network that fuses them to detect and locate objects. Different projections expose different structure; combining them recovers more than any single view. The CPC tags read like an autonomy bill of materials: G01S 17/931 (vehicle LiDAR), B60W 60/0011 (autonomous control), G06N 3/045 (neural networks), G06V 20/58 (detecting road objects).

“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

Here's why this is the arms-dealer move and not a robotaxi move. A company like Waymo or Zoox patents perception to power its own fleet. Nvidia patents perception that is deliberately platform-agnostic — it runs on Nvidia silicon and slots into anyone's vehicle. The same horizontal logic that made Nvidia the default supplier of training compute is being extended to the autonomy software layer: own the components, sell to all the assemblers.

Vision-only versus mapped is a cost bet — and Nvidia is hedging it by supplying the LiDAR-perception side to whoever wants it. That's a tell about where it thinks the durable margins are: not in declaring a winner between sensor camps, but in being the indispensable supplier regardless of which camp wins. The boring incumbent it threatens isn't a robotaxi company; it's every autonomy team's in-house perception group.

What the patent doesn't claim is field performance — detection rates, false-positive behavior, latency under load. A method for fusing views is not a benchmark, and the safety-relevant numbers live in validation data the grant never touches. A licensable perception block still has to be validated inside each customer's safety case.

For readers tracking who actually profits from autonomy, the LiDAR-perception patent is the quiet evidence. Nvidia is filing the horizontal IP — the parts every stack needs — which is how an arms dealer wins a war it never enters.