Here is what actually issued. On June 16, 2026, the U.S. Patent and Trademark Office granted NVIDIA Corporation U.S. Patent No. 12,657,819, titled “Surface estimation.” This is a grant, not a published application — the kind code is B2, and the rights are now live. The core invention is a training pipeline: a way to teach a deep neural network to estimate the 3D structure of a road surface without ever sending a sensor-equipped vehicle out to gather and hand-label the data the network learns from.
The independent claim language centers on the data-generation step, and that is where the substance lives. Rather than collecting real drives, the method generates a variety of synthetic 3D road surfaces by modeling each one parametrically — varying the parameters to simulate changes in road direction and lateral surface slope. A synthetic surface is built by modeling a longitudinal 3D curve and expanding that curve into a full 3D surface, which is then sampled to form a ground-truth projection image, in practice a 2D height map. To produce the matching input the network will see at inference time, the method applies a known pattern of which pixels would remain unobserved during 3D structure estimation, turning a dense ground-truth image into a sparse projection image. The DNN trains on the sparse-input, dense-output pairs.
"In various examples, to support training a deep neural network (DNN) to predict a dense representation of a 3D surface structure of interest, a training dataset is generated using a parametric mathematical modeling."— U.S. Patent No. 12657819, source
The named inventors are Kang Wang, Yue Wu, Minwoo Park, and Gang Pan. The CPC classification tells you where this sits: the lead symbol is G06T 17/05 (geometric modeling for geographic or topographic features), and the training-data angle is captured by G06F 18/214 (generating training data for pattern recognition). The remaining symbols — a cluster under B60W (B60W 30/09, 30/143, 40/06, 50/06, 60/001) and G06V 20/05 and G06V 20/58 — anchor the application squarely in the driver-assistance and autonomous-driving stack. This is a perception patent, classified less as abstract machine learning (you will not find a primary G06N 3 symbol here) and more as the applied geometry of how a vehicle reconstructs the road ahead.
Why generate the road instead of driving it
The deflationary read on “synthetic training data” is that it is everywhere now, and it is. But the claim is specific about how the synthetic data is constructed, and that specificity is the point. Two problems dog real-world 3D surface ground truth. First, dense, accurate height maps of road surfaces are expensive to acquire and harder to label — you need high-end sensing, careful calibration, and a lot of road. Second, real sensors do not return dense surfaces; they return sparse, partial observations, full of gaps where the geometry was occluded or simply not measured. A network trained naively on clean dense data learns to expect input it will never actually receive.
The patented method addresses both at once. Because the synthetic surface starts from a parametric curve, the ground truth is exact and free — you know the surface because you defined it. And because the method then applies a known masking pattern to simulate which pixels go unobserved, the network is trained on inputs that mimic the sparseness of a real sensor while still being graded against a perfect dense target. That is the structural trick: perfect supervision paired with realistically degraded input. Varying road direction and lateral slope across the generated set is what gives the network the diversity it needs to generalize to roads it was never shown.
What the grant covers, and what it does not
Read the claim, then we will talk scope. What issued is a training-and-data-generation methodology tied to 3D road-surface estimation, not a monopoly on synthetic data or on deep networks generally. The novelty NVIDIA is asserting lives in the chain — parametric curve to expanded 3D surface to sampled height-map ground truth to masked sparse input to trained DNN — and the value of the grant is bounded by how tightly that chain is recited in the independent claim. The abstract describes “various examples” and “an example embodiment,” standard spec hedging, but the enforceable surface is the claim language, and the claim language is built around generating the dataset by parametric modeling.
For NVIDIA, the strategic fit is obvious. The company sells the compute and the software platforms that autonomous-driving developers train and run perception models on, and a patent that covers a concrete, cost-saving way to manufacture road-surface training data slots neatly into that portfolio. It is the kind of applied-perception grant — useful, narrow, well-classified — that accretes into a defensible estate rather than a single headline. Granted, not just filed: as of June 16, this one is enforceable, and it is one more brick in NVIDIA's autonomous-driving wall.