Enabling claim or aspiration? That is the first question to ask of any autonomy patent, and NVIDIA's grant US12651465B2, "Multi-view deep neural network for LiDAR perception" (issued June 9, 2026), answers it on the face of its classifications. The CPC list includes B60W 60/0011, B60W 60/0016, and B60W 60/0027 — autonomous driving control — alongside G01S 17/931 and G01S 17/89, the codes for LiDAR systems mounted on vehicles. This is not a method floating free of a vehicle; it is claimed inside the driving stack.
The mechanism in the title is the load-bearing part. "Multi-view" means the system does not feed raw LiDAR returns into a single network and hope. It projects the point cloud into more than one representation — typically a perspective (range-image) view that preserves sensor geometry, and a bird's-eye or top-down view that preserves metric layout — and processes those views together. Each view is good at something the other is bad at: perspective views keep fine angular detail, top-down views keep distances and footprints clean. Fusing them is how you get both.
Read the inventor list and the intent is unambiguous. The named inventors include David Nister and colleagues long associated with NVIDIA's autonomous-driving research — the people who build the perception layer that runs on the company's automotive compute. A patent is a method, not a shipped feature, but the assignee, the inventor roster, and the B60W classification together tell you this method is meant to run on a car, not in a lab demo.
What the grant does not claim is just as important for an IP reader. It covers a multi-view neural-network approach to LiDAR perception; it does not claim LiDAR itself, nor object detection in general, nor any particular sensor hardware. The scope is the architecture — how views are formed and combined — and the dependent claims narrow it further. Anyone reading this as "NVIDIA owns LiDAR perception" is overreading the abstract; the defensible position is the specific multi-view processing pipeline.
Why it matters for the autonomy IP race: LiDAR perception is the contested ground between camera-first players and sensor-fusion players, and the network architecture that turns raw returns into objects is where the genuine engineering — and the genuine claims — live. NVIDIA does not sell cars, but it sells the silicon and the software stack that other companies' cars run, and a granted perception method is a piece of that stack it now owns outright.
For readers tracking the B60W class, this grant is a marker. The autonomy-control CPC bucket is filling with sensor-perception methods like this one, and the assignees are increasingly the compute and software suppliers rather than the automakers themselves. The roadmap tell is in who holds the perception IP — and here, it is the chip company.