Nvidia Patents a Way to Plan Robot Paths on a GPU in Parallel
Planning a route for a self-driving car or robot arm sounds simple until you realize the machine has to weigh thousands of possible moves every fraction of a second. Nvidia's new patent tackles that bottleneck by offloading the whole calculation to a GPU, running every possible path at the same time.
How Nvidia's path planner picks the best route for a robot
Imagine you're driving and need to parallel park. Your brain quickly rules out bad angles, picks the best approach, and commits. An autonomous vehicle has to do the same thing, but mathematically, evaluating hundreds of possible steering angles and speeds almost simultaneously.
Nvidia's patent describes a system that pushes that evaluation onto a GPU (the same kind of chip that renders video game graphics), because GPUs are built to do many calculations at once rather than one at a time. Instead of checking each possible path one by one, the system checks them all in parallel, then picks whichever path costs the least (factoring in things like how sharp a turn is, or whether switching maneuvers mid-route carries a penalty).
The result is that an autonomous machine, whether a self-driving car, a warehouse robot, or a delivery drone, can figure out where to go and how to get there much faster than older, single-threaded approaches would allow.
compute a first candidate cost for the autonomous or semi-autonomous machine to reach a common pose in a discretized space of poses using a first maneuver based at least on applying, to a first cost for the autonomous or semi-autonomous machine to reach a first pose of the first maneuver …
Translation: The robot calculates the cost of different movement options to reach a specific position and orientation.
How the GPU evaluates hundreds of maneuvers simultaneously
The patent describes a path planning algorithm for autonomous or semi-autonomous machines. The core problem it solves is: given a map of where the machine is and where it needs to go, how do you quickly find the best sequence of maneuvers?
The system works in a discretized pose space (a grid of possible positions and orientations the machine could be in) and evaluates multiple candidate trajectories at the same time. A key insight is that many turns with the same radius and starting angle are just shifted copies of each other along the orientation axis, so they can be processed together on a GPU rather than computed separately.
For each candidate path, the system computes a candidate cost, which combines:
- The cost to reach a particular position along that maneuver
- A transition penalty for switching from one maneuver type to another (for example, going from straight driving to a left turn)
All these costs are encoded as bit vectors (compact binary representations), and GPU threads process them in synchronization using a technique called parallel reduction (a standard GPU pattern for combining many results into one while respecting ordering dependencies). The path with the lowest total cost wins.
Turns or turn portions that have the same turn radius and initial orientation can be translatively shifted along and processed in parallel along the orientation axis as they are translated copies of each other, but with different starting points.
Translation: The system speeds up path planning by processing identical turn shapes simultaneously instead of one by one.
What faster path planning means for self-driving and robotics
Path planning is one of the most computationally expensive parts of running an autonomous vehicle or robot. If the planner is slow, the machine either has to wait before acting or work with a less thorough search, both of which hurt safety and performance. Shifting this workload onto a GPU could let vehicles react faster and consider more options in the same time budget.
Nvidia already sells the DRIVE platform for autonomous vehicles and the Isaac platform for robotics, and both rely heavily on GPU compute. A patent like this fits directly into that stack, suggesting Nvidia is working to make its hardware-software combination faster at one of the hardest real-time tasks in autonomy.
This is practical, infrastructure-level engineering rather than a flashy AI announcement, but it matters. Path planning speed is a genuine ceiling on how capable autonomous systems can be, and a GPU-native approach that exploits parallel structure in turn geometry is exactly the kind of work that ends up shipping in real products. Worth paying attention to if you follow self-driving or robotics.
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
21 drawing sheets from US 2026/0227797 A1 · click any drawing to enlarge
Show the other 13 sheets
Want this weekly breakdown for a company we don't cover? Patentlyze Pro →