Nvidia Files Patent for Helping Robots Figure Out Where They Are Indoors
Switch a robot on in the middle of a warehouse and it has no idea where it is. Nvidia's filing has it look through two camera lenses, match what it sees to a saved map, and work out its position.
What Nvidia's two-camera robot locator does
Every time a warehouse robot is switched on or carried to a new spot, it faces the same question: where am I? GPS doesn't help much inside a building, so the robot has to work it out by looking around.
Nvidia's patent application has the robot snap a photo with each of its two camera lenses, like a pair of eyes. It compares both photos against a stored map of earlier photos, each tagged with where it was taken. Checking how each view lines up with the map photo, and using the known gap between the two lenses, the robot can work out which way it is facing and how far it is from where the map photo was shot.
The filing says this skips an older step that guesses at 3D landmarks, which can introduce errors. Instead it works from the flat images and then fine-tunes the answer.
… determining at least a first relative pose between the left image and the map image and a second relative pose between the right image and the map image; determining, based at least on the first relative pose and the second relative pose, a localized pose associated with the machine within the environment; …
Translation: The system compares both camera views to the map to figure out the robot exact location.
How two lenses fix the distance guesswork
The system starts with a stereo camera (two lenses mounted a fixed distance apart, like eyes) on a machine such as a robot or vehicle. It first finds a stored map image that looks like what the camera sees. Then it compares the left image and the right image against that map image.
From each comparison, it builds an essential matrix (a small grid of numbers describing how two camera views relate) and breaks it down into a relative pose, meaning a rotation plus a direction of movement. The catch is that each comparison alone gives direction but not true distance. The filing calls this a translation scale problem.
To fix it, the system adds a third relative pose: the left-to-right relationship, known exactly from the camera's calibration. Because that spacing is fixed, simple linear fitting (basic line-fitting math) pins down the real distance for both comparisons.
Optionally, it then refines the answer by shrinking the gap between matched points and epipolar lines (lines in one image where a matching point from another view must fall). The refined pose, combined with the known position of the map image, gives the machine's final location and orientation. The machine then acts on it, for example by navigating without hitting things.
Additionally, the predicted pose may be refined using one or more optimization techniques, such as by minimizing distances between feature points of the images and epipolar lines associated with the points.
Translation: The rough location estimate is then fine tuned using geometry math to make it more precise.
What this means for warehouse robots
If you've watched a robot vacuum get confused after being picked up and moved, you know the problem. Starting up or relocating indoors, where GPS signals struggle, is a weak spot for any machine. A faster, more accurate way to get its bearings matters in the warehouses, stores and offices the filing names as example settings.
The filing says working from flat 2D images, rather than estimated 3D landmarks, avoids a source of errors, and that the refinement step adds accuracy. It also describes testing the method in simulation (Nvidia's DriveSIM and Isaac Sim) before real-world use. The claims are written broadly, covering vehicles, robots and other machines, so this is a building block for navigation rather than a single product.
Nvidia's 51st filing in our robotics coverage since May adds to a run that includes one on maps that update in place and one on vision that redraws objects.
No new cameras required. The method described here runs on a stereo camera, which is simply two lenses mounted side by side, a setup many robots and automated vehicles already carry. That makes this a software update, not a hardware purchase.
Before any machine can use it, someone has to build the map: drive or walk a space, collect photos at known positions, and keep that library current. That upfront work makes warehouses and other controlled, repeatedly-visited spaces the natural first home for this approach.
For a factory fleet, the economics are straightforward because one map serves every machine on the floor. For a consumer product, the mapping burden is substantial and has to be solved separately before this technique can ship as anything a regular person would touch.
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
25 drawing sheets from US 2026/0320822 A1 · click any drawing to enlarge
Show the other 17 sheets
Want this weekly breakdown for a company we don't cover? Patentlyze Pro →
Be the first to weigh in