Nvidia Patents a Radar Method for Tracking How Fast Objects Are Moving Around Autonomous Machines
A newly published Nvidia patent tackles one of the quiet headaches in autonomous driving: figuring out exactly how fast something near your vehicle is moving, even when your sensors only catch a slice of the picture.
How Nvidia's radar system figures out object speed
Ever watched a car merge onto the highway and wondered how fast it was actually going? A self-driving system has to answer that question constantly, in fractions of a second, using sensor data that is often incomplete.
Nvidia's patent describes a system that estimates an object's speed by combining two radar readings: how fast the object appears to be moving toward or away from the sensor (called range rate), and what angle that object sits at relative to the vehicle. On their own, neither reading tells the full story. But the system also factors in what it already calculated about that object a moment ago, using that previous estimate as a reference to sanity-check the new readings.
The result is a continuously updated picture of what's around an autonomous machine, whether that's a self-driving car, a robot, or a drone, and how fast each nearby object is moving. That running estimate feeds directly into how the machine decides to steer, brake, or hold course.
… the machine performs one or more control operations based at least on a velocity estimate corresponding to an object and determined based at least on: a range rate measurement corresponding to a portion of the object; and an angle measurement corresponding to the portion of the object.
Translation: The vehicle steers or brakes by calculating how fast and where nearby objects are moving.
How range rate, angle, and prior estimates combine
The patent centers on a velocity estimation pipeline for machines equipped with radar sensors. Radar can measure the range rate of an object, which is the speed at which it is closing in on or moving away from the sensor along a straight line. But that single number doesn't tell you whether something is crossing in front of you, moving in parallel, or heading straight for you.
To fill that gap, the system also reads the angle at which the object sits relative to the sensor. By combining that angle with a previously calculated velocity estimate for the same object, the system can predict what the range rate should be if that earlier estimate was accurate. It then compares that predicted number with the actual measured range rate. The gap between the two tells the system how much to revise its estimate of the object's true speed and direction.
This iterative comparison runs on systems-on-a-chip (SoCs), which are single chips that pack a CPU, GPU, and hardware accelerators together. The SoCs process incoming sensor data fast enough to keep the velocity estimates current as the machine moves and the scene changes.
- Range rate measurement: how fast the object is approaching or receding
- Angle measurement: where the object sits relative to the sensor
- Prior state estimate: the system's best previous guess at object speed
- Expected vs. Measured comparison: the mechanism that refines the estimate each cycle
… an expected range rate for the object may be determined using an angular measurement and a previous velocity estimate of the object. The expected range rate may be compared with the range rate measurement to determine a velocity estimate for the object.
Translation: It predicts how fast an object should be moving based on past data and checks it against current radar readings.
What sharper object tracking means for self-driving safety
For you as a passenger or pedestrian near an autonomous vehicle, this kind of system is part of what keeps a self-driving machine from misjudging a car that's drifting into its lane or a cyclist who is accelerating through an intersection. A bad velocity estimate leads to a bad braking decision. Getting it right, continuously and quickly, is not optional.
Nvidia's run of autonomous-machine sensor filings reflects how much the company is investing in the compute and perception layers that make self-driving trustworthy. This patent sits at the intersection of radar signal processing and the onboard chips Nvidia already sells to automakers, suggesting the technology is intended to work with hardware already in the pipeline rather than requiring entirely new sensors.
Nvidia's 53rd filing we've tracked since May on our self-driving sensing race follows one on detecting hidden objects and one on rebuilding occluded shapes, keeping the company deep in what sensors cannot directly see.
Radar already sits on most self-driving vehicles, but pulling accurate speed data from it has always required careful math. This patent describes a practical check: compare what the radar just measured against what the system already predicted, then refine the estimate. The driver never sees that loop running, but they'd feel its absence as an unnecessary hard brake or a missed warning.
The improvement is steady rather than dramatic. The real contribution is in how this check is wired into Nvidia's specific chip architecture, making the calculation fast enough to matter in real traffic.
For a passenger, that translates to a system that handles a cyclist cutting across an intersection or a car merging at speed with less hesitation and fewer false alarms. Small accuracy gains in object tracking accumulate into a vehicle that earns trust over miles rather than losing it.
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 drawing sheets from US 2026/0259313 A1 · click any drawing to enlarge
Want this weekly breakdown for a company we don't cover? Patentlyze Pro →