Get the newsletter
Nvidia · AI / ML · Filed May 13, 2025 · USPTO published Nov 19, 2026 · This week's release

Nvidia Patents Object Outlines That Trade Detail for Speed on Cars and Robots

Most vision software outlines a stop sign and a bicycle with the same fixed budget of points. Nvidia's patent application has the AI choose how many corner points each object needs, and lets engineers turn the overall detail up or down to match the chip.

Autonomous vehicles and mobile robots equipped with cameras, radar, and lidar sensors to map nearby objects in real time. Drawing from patent filing US 2026/0365197 A1.
Autonomous vehicles and mobile robots equipped with cameras, radar, and lidar sensors to map nearby objects in real time.All 12 drawings ↓
Publication number
US 2026/0365197 A1(US20260365197A1)
Applicant
NVIDIA Corporation
Filed · Published
May 13, 2025 · Nov 19, 2026
Inventors
Rui SHEN
Status when we published
Approved; patent expected soon
Document
20 claims
US classification
701/23
Examiner
LEE, TYLER J (Art Unit 3663)

What Nvidia's adjustable object outlines do

Today, most systems that outline an object in a photo work from a fixed budget: a box with set corners, a curve with a set number of control points, or a grid of pixels marked yes or no. A bicycle needs far more detail than a stop sign, yet the budget stays the same. Older wave-based methods also tend to round off sharp corners.

Nvidia's filing describes a different recipe. Its AI describes an outline as a blend of smooth waves, and it also decides for itself how many corner points a given object needs. A stop sign might need a few. An oval might need none.

The useful twist is that you can turn the detail up or down. A small chip inside a car or robot can use a quick, rough outline, while a data center can afford a much finer one, all from the same basic method.

From the filing · CLAIM 1
… predict, using a machine learning model, a quantity of time-domain keypoints; calculate a set of time-domain keypoints based at least on the predicted quantity and a set of Fourier coefficients associated with a Fourier descriptor …

Translation: An artificial intelligence model figures out how many points are needed to outline an object.

How Nvidia's AI rebuilds an outline from waves

The system starts with an ordinary 2D camera image that shows an object. A trained neural network (a program that learns patterns from examples) looks at it and produces two things: a set of Fourier coefficients and a keypoint count. Fourier coefficients are numbers that describe a closed shape as a stack of smooth waves added together. More waves mean finer detail.

A later math step, an inverse Fourier transform (the reverse of breaking a shape into waves), turns those numbers back into actual positions on the image. The predicted keypoint count decides how many of those positions are marked as key corners. If it predicts zero, as it might for a circle or oval, the system instead samples a preset number of points, such as 25, 50 or 100, along the curve.

The detail dial is the degree of the descriptor, called K, which sets how many coefficients (2K+1) the network predicts. The filing gives K=7 as a fit for a vehicle or robot with limited computing power, and K=100 to 200 for servers. A multi-exit network can offer both: earlier layers output a K=7 outline, later layers a K=200 one.

Training uses images paired with correct outlines, and the model is adjusted until its outlines match. The claims end with a machine acting on the result through control or navigation operations. One claim describes a vehicle with CPUs, GPUs, accelerators and outward-facing sensors.

From the filing · THE ABSTRACT
By varying the number of predicted Fourier coefficients, the disclosed techniques are operable to generate boundary representations having differing degrees of accuracy and requiring differing amounts of computational resources.

Translation: Adjusting the detail level lets the system trade precision for processing speed.

Why a sliding detail dial matters for robots

For you, the practical effect would show up in machines that have to see edges quickly: driver-assist systems, warehouse robots, delivery bots. A car cannot wait on a cloud server to trace a pedestrian, so a fast, cheap version matters. The same method can produce a finer version where computing power is plentiful, such as when automatically labeling data used to train other AI. The filing also lists uses in simulation, augmented reality and security cameras.

The patent office has approved this application, and once Nvidia pays the final fee it should be issued as a granted patent. A patent is not a product announcement. The filing names Nvidia's DriveSIM and Isaac Sim tools only as places the method could be tested.

That makes this Nvidia's 117th filing we've tracked since May in the self-driving sensing race, following one on a location double-check system and one on a radar chip that pre-processes data.

Source: US 2026/0365197 A1 at the USPTO (PDF). Full patent text and figures are in the official publication.

Get our take in your Top Stories

Liked this breakdown? Add Patentlyze as a preferred source on Google, and our plain-English take shows up more often in your Top Stories the next time Nvidia patent news breaks.

There are more where this came from

We read every patent application Big Tech publishes and send you the ones worth knowing. Plain English, free, every week.

The drawings · 12 sheets

12 drawing sheets from US 2026/0365197 A1 · click any drawing to enlarge

Computer hardware that trains and runs the vision AI, drawing sheet 1 from US 2026/0365197 A1
Sheet 1 Computer hardware that trains and runs the vision AI
Parts of the system that teach the AI object shapes, drawing sheet 2 from US 2026/0365197 A1
Sheet 2 Parts of the system that teach the AI object shapes
Steps for training the AI to outline shapes efficiently, drawing sheet 3 from US 2026/0365197 A1
Sheet 3 Steps for training the AI to outline shapes efficiently
How the trained AI turns an image into an outline, drawing sheet 4 from US 2026/0365197 A1
Sheet 4 How the trained AI turns an image into an outline
Steps for generating quick object outlines from photos, drawing sheet 5 from US 2026/0365197 A1
Sheet 5 Steps for generating quick object outlines from photos
Where cameras and radar mount on cars and robots, drawing sheet 6 from US 2026/0365197 A1
Sheet 6 Where cameras and radar mount on cars and robots
Locations of driving controls and sensors across a car, drawing sheet 7 from US 2026/0365197 A1
Sheet 7 Locations of driving controls and sensors across a car
Electronic controls linking sensors to vehicle steering and brakes, drawing sheet 8 from US 2026/0365197 A1
Sheet 8 Electronic controls linking sensors to vehicle steering and brakes
Show the other 4 sheets
Computer chips packed into a single vehicle processor, drawing sheet 9 from US 2026/0365197 A1
Sheet 9 Computer chips packed into a single vehicle processor
How cloud servers send updated navigation data to a car, drawing sheet 10 from US 2026/0365197 A1
Sheet 10 How cloud servers send updated navigation data to a car
Training self driving AI in data centers for cars, drawing sheet 11 from US 2026/0365197 A1
Sheet 11 Training self driving AI in data centers for cars
Computer hardware parts that run the vision software, drawing sheet 12 from US 2026/0365197 A1
Sheet 12 Computer hardware parts that run the vision software
Patent filing page

Watchlists this filing belongs to

All watchlists →
Reader comments

No comments yet

Start the discussion

Real name or a handle, either is fine. Comments are read by a person before they appear, so allow a little time. Keep it about the filing.