Nvidia Files Patent for Robots That Handle Surprises They Were Never Trained On
Instead of reacting to raw camera pixels, this Nvidia patent has a robot first write down what is true about its surroundings, then pick its next move from that list.

- Publication number
- US 2026/0364786 A1(US20260364786A1)
- Applicant
- NVIDIA Corporation
- Filed · Published
- Feb 12, 2026 · Nov 19, 2026
- Inventors
- Christopher Jason Paxton, Kei Kase, Dieter Fox
- Status when we published
- Waiting for an examiner
- Document
- 20 claims
- US classification
- 701/27
- Parent application
- Continuation of 16/791,943 (filed Feb 14, 2020)(16791943)
What Nvidia's robot checklist planner does
Every time something in a robot's workspace shifts, a cup gets nudged or a lid gets closed, the robot has to work out what changed and what to do about it. Many robots handle this badly. They react to raw camera images, and anything they were not trained on can leave them slow or stuck.
Nvidia's patent splits the job in two. First, a neural network (software that learns from examples) looks at the camera feed and writes a short checklist of facts about the scene. Then a planner works backward from the goal to see which conditions have to be true first, and picks steps that make them true.
Because the robot rechecks the facts after every move, a surprise just changes the checklist, and the plan adjusts with it.
… select one or more symbolic operators that, if applied to the one or more logical predicates, would cause the one or more preconditions to be satisfied; and cause the device to perform one or more actions determined based at least on the one or more symbolic operators.
Translation: The system picks the right logical steps to solve the problem and makes the robot carry them out.
How the robot works backward from its goal
The first independent claim covers a processor that starts from a goal and identifies the preconditions (things that must be true before the goal can be reached). Other claims say these are backpropagated from the goal state, meaning worked backward from the end result so each step sets up the one after it.
One or more neural networks (software trained on examples) turn sensor data, such as camera images, into logical predicates, which are symbolic statements describing the environment. The system then selects symbolic operators, named actions a robot can perform, that would satisfy the preconditions, and the device acts on them.
The description lays out a two-level design:
- A high-level policy reads a few recent camera frames plus the robot's joint readings and infers the current logical state.
- A low-level policy, another neural network, turns the current operator into the joint positions the robot should move to.
- A loop repeats: check whether the goal is met, read the sensors, infer the state, pick an operator, act.
The predicates are recomputed from current sensor data, so the plan follows what the robot sees now. During training, the networks also predict depth and object-outline images as side tasks to help them learn spatial relationships.
Apparatuses, systems, and techniques to facilitate robotic execution using neural networks to perform complex, multi-step tasks in situations for which a robot has not been trained.
Translation: This technology helps robots use AI to handle difficult new situations they never saw before.
What this means for robots facing surprises
If you have watched a robot freeze because something sat an inch from where it expected, you have seen the problem here. Robots that only react to camera images struggle when a situation is new, and hand-built rule trees get huge as more situations are added. A planning layer that reasons about what must be true next is a way to cope with surprises without writing a rule for each one.
For anyone hoping for robots in homes, warehouses or hospitals, multi-step jobs in messy surroundings are the hard part. The filing describes a robot arm with cameras and joint sensors, and says the aim is tasks the robot has not been trained for. It is a patent application, so it shows a direction of research, not a product on a shelf.
That makes this Nvidia's 39th filing we've tracked since May in our robot grasping and movement work, adding to earlier applications like teaching by watching hands and replanning when objects shift.
Source: US 2026/0364786 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.
We read every patent application Big Tech publishes and send you the ones worth knowing. Plain English, free, every week.
The drawings · 59 sheets
59 drawing sheets from US 2026/0364786 A1 · click any drawing to enlarge
Show the other 51 sheets

No comments yet