Get the newsletter
Toyota · AI / ML · Filed Apr 6, 2026 · USPTO published Nov 19, 2026 · This week's release

Toyota Patents a System That Tells Robots When a Task Is Going Wrong

A robot that can follow a spoken order still has no idea when it is about to drop the pot. Toyota's research arm has filed a patent application for a watcher that reads the robot AI's own internal signals and flags trouble early.

Camera feeds track a robotic arm as an AI flags the exact moment an attempt to move a coffee pot goes wrong. Drawing from patent filing US 2026/0369045 A1.
Camera feeds track a robotic arm as an AI flags the exact moment an attempt to move a coffee pot goes wrong.All 18 drawings ↓
Publication number
US 2026/0369045 A1(US20260369045A1)
Applicant
Toyota Research Institute, Inc.
Filed · Published
Apr 6, 2026 · Nov 19, 2026
Inventors
Qiao Gu, Yuanliang Ju, Shengxiang Sun, Igor Gilitschenski, Haruki Nishimura, Mikhal ltkina, Florian Shkurti
Status when we published
Waiting for an examiner
Document
20 claims
US classification
382/155
Parent application
Claims priority from a provisional application 63806272 (filed May 15, 2025)

What Toyota's robot failure detector does

Ever watched a robot vacuum push at the same chair leg over and over, totally unaware it's stuck? Machines that follow instructions are often bad at noticing when things are going wrong. That matters a lot more when the robot is an arm carrying a hot pot.

Toyota Research Institute's patent application describes a watcher that sits beside a robot's AI brain. The AI takes in camera images and a written instruction like "put the pot on the stove," then produces movements. The watcher peeks at the AI's internal numbers as it works and boils them down to a single score: how likely is this going wrong?

If that score crosses a limit, the robot can stop, back up, or send you an alert. The limit isn't fixed. It shifts as the task goes on, and it is set using examples of the robot doing things right.

From the filing · CLAIM 1
extract features from one or more layers of a vision-language-action (VLA) model; predict a failure score by processing the features; determine a time-varying threshold for detecting a failure by the VLA model …

Translation: The system analyzes the robot brain's inner workings to calculate a score for potential errors over time.

How the score and moving threshold work

Toyota's filing covers a failure detector that attaches to a vision-language-action (VLA) model, an AI that takes camera images plus a written instruction and outputs robot movements. The detector works in three steps:

  • Read the features. It pulls the model's internal numbers (called features) from the last layer, just before they turn into movements.
  • Score the risk. A small neural network, either an MLP (a simple stack of layers) or an LSTM (a network that remembers earlier steps), turns those numbers into one failure score that updates as the task goes on.
  • Compare to a moving limit. The limit comes from functional conformal prediction (a statistical method) applied to a held-out set of successful runs, so it changes over time. At the setting the filing favors, false alarms on successful runs are meant to be capped at about 15 percent.

When the score crosses the limit, the system can send a control signal to stop or back up, or send an alert to a user device.

The filing says the detector was trained on some tasks and tested on tasks it had never seen, using three existing robot AI models in simulation and on real robots. It adds about 0.73 milliseconds, against 149 milliseconds for the underlying model.

From the filing · THE ABSTRACT
… detect the failure based on the failure score exceeding the time-varying threshold; and perform an action in response to detecting the failure.

Translation: It flags a mistake when the error score gets too high and then takes steps to fix the problem.

Why robots need to know when they're failing

If robots are going to work around people, they need a way to say "this isn't going well" before something breaks or spills. This detector aims to give robot AI that habit without slowing it down, and without needing a fresh round of training for every new chore. Because the claims also mention vehicle control, the idea isn't limited to robot arms.

The limits are worth knowing. The filing admits the detector needs hundreds of success and failure examples from other tasks before it can work, and on varied real-world tasks its accuracy was modest. It also only raises the alarm. It doesn't make the robot any better at the job.

Source: US 2026/0369045 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 Toyota patent news breaks.

There are more where this came from

We read the new patents from Big Tech, carmakers and more and send you the ones worth knowing. Plain English, free, every week.

The drawings · 18 sheets

18 drawing sheets from US 2026/0369045 A1 · click any drawing to enlarge

Mapping robot mistakes into a shared failure zone, drawing sheet 1 from US 2026/0369045 A1
Sheet 1 Mapping robot mistakes into a shared failure zone
Steps for scoring robot actions and flagging errors, drawing sheet 2 from US 2026/0369045 A1
Sheet 2 Steps for scoring robot actions and flagging errors
Test accuracy scores across simulated robot tasks, drawing sheet 3 from US 2026/0369045 A1
Sheet 3 Test accuracy scores across simulated robot tasks
Real world test results alongside everyday test chore commands, drawing sheet 4 from US 2026/0369045 A1
Sheet 4 Real world test results alongside everyday test chore commands
Success rates and simulated chores for test robots, drawing sheet 5 from US 2026/0369045 A1
Sheet 5 Success rates and simulated chores for test robots
Benchmark dataset sizes and fine tuning settings, drawing sheet 6 from US 2026/0369045 A1
Sheet 6 Benchmark dataset sizes and fine tuning settings
Parameters configured for each failure detection technique, drawing sheet 7 from US 2026/0369045 A1
Sheet 7 Parameters configured for each failure detection technique
Additional tuning settings used across test methods, drawing sheet 8 from US 2026/0369045 A1
Sheet 8 Additional tuning settings used across test methods
Show the other 10 sheets
Parameter options tested across error detection methods, drawing sheet 9 from US 2026/0369045 A1
Sheet 9 Parameter options tested across error detection methods
Additional tuning settings tested for robotic mistake tracking, drawing sheet 10 from US 2026/0369045 A1
Sheet 10 Additional tuning settings tested for robotic mistake tracking
Further parameter variations evaluated for task error monitors, drawing sheet 11 from US 2026/0369045 A1
Sheet 11 Further parameter variations evaluated for task error monitors
Model configurations used in larger robot test setups, drawing sheet 12 from US 2026/0369045 A1
Sheet 12 Model configurations used in larger robot test setups
Continuation of test configurations for evaluating task monitoring, drawing sheet 13 from US 2026/0369045 A1
Sheet 13 Continuation of test configurations for evaluating task monitoring
Experimental results comparing mistake detection accuracy, drawing sheet 14 from US 2026/0369045 A1
Sheet 14 Experimental results comparing mistake detection accuracy
Accuracy scores across different training limits and AI backbones, drawing sheet 15 from US 2026/0369045 A1
Sheet 15 Accuracy scores across different training limits and AI backbones
Parts that keep a robot from making mistakes, drawing sheet 16 from US 2026/0369045 A1
Sheet 16 Parts that keep a robot from making mistakes
Steps the robot follows to spot mistakes and step in, drawing sheet 17 from US 2026/0369045 A1
Sheet 17 Steps the robot follows to spot mistakes and step in
Computer parts that power the robot mistake detector, drawing sheet 18 from US 2026/0369045 A1
Sheet 18 Computer parts that power the robot mistake detector
Patent filing page

Watchlists this filing belongs to

All watchlists →
Reader comments

No comments yet

Start the discussion

Real name or a handle, either is fine. Comments are read by a person before they appear, so allow a little time. Keep it about the filing.