The patent (US 2026/0115906 A1, filed July 11, 2025) describes a computer-implemented method built around two distinct ML models working in sequence: generation first, then filtering.

The two-stage idea

Imagine you're teaching a robot arm to pick up a coffee mug. The robot needs to figure out not just where to grab it, but how — the angle, the grip, the approach. Get any of those wrong and the mug ends up on the floor.

Nvidia's patent describes a two-step AI approach. First, one model looks at sensor data (think cameras or depth sensors) and generates a bunch of candidate grip positions — essentially brainstorming a list of ways the robot could grab the object. Then a second model acts like a quality filter, reviewing that list and picking out only the grips that are likely to actually work. The surviving grips get assembled into a final grasping plan, and the robot executes it.

By splitting "generate ideas" and "pick the best idea" into two separate trained models, the system can be smarter and more reliable than a single do-everything model.

Inside the pipeline

The patent describes a computer-implemented method built around two distinct machine learning models in sequence.

Stage 1 — Pose generation. A first trained ML model takes in raw sensor data (point clouds from depth cameras or similar 3D input) and outputs a set of candidate grasp poses — specific combinations of position and orientation for the robot's hand or gripper.

Stage 2 — Pose filtering. A filtering model evaluates those candidate poses and selects a refined subset, removing unlikely or physically infeasible options.

Planning and execution. The filtered poses feed into a grasping plan, which the robot then executes.

The key architectural insight is decomposition: rather than asking one monolithic model to go from pixels to action, Nvidia's system breaks the task into generation and selection — a pattern borrowed from fields like image synthesis (think diffusion models' noise-then-denoise logic). It makes each sub-model easier to train and evaluate independently.

The inventors — several of whom are well-known robotics researchers (Dieter Fox alone has co-authored some of the most-cited work in robot perception) — appear to be formalizing techniques from Nvidia's Isaac robotics platform into patentable method claims.

Why this matters

Robot manipulation is the bottleneck holding back practical automation in warehouses, factories, and homes. Most current systems are brittle: they work for a narrow set of known objects under controlled lighting, and fall apart when conditions change. A learned two-stage pipeline that generalizes across arbitrary objects from sensor data alone is exactly the kind of capability that closes that gap.

For Nvidia, this fits squarely into its Isaac robotics push — the company wants to be the platform layer that robot makers build on, the same way it became the platform layer for AI training. If Nvidia can patent and productize core manipulation algorithms, it strengthens the moat around that platform considerably.

Editorial take

Focused, technically credible filing from a team with serious robotics credentials. The two-stage generate-then-filter architecture is a well-motivated design choice, not just patent-filling. Worth watching because it signals Nvidia is thinking hard about making manipulation a software product, not just a research problem.

Originally published at patentlyze.com — plain-English breakdowns of every Big Tech patent at the USPTO.