What is physical AI? Factory news from November 2026

For the past two years, most of us have pictured AI as a chat tool that writes text on a screen. In the first weeks of November 2026, however, news from industry shared a common theme: physical AI. That means AI that does not only talk, but sees, understands and does real, hands-on work.

In this article we explain physical AI in plain language, read three real news stories from this month through the eyes of a factory manager, and discuss which steps make sense for your own business.

What does physical AI mean?

A classic industrial robot repeats the motion it was taught, at the same point, in the same way. If a part shifts by a few centimetres or the lighting changes, it often stops or makes a mistake. That is why robot installations take a long time and every product change calls for reprogramming.

With physical AI, the robot sees its surroundings through a camera, interprets what it sees and adjusts its motion accordingly. Three pieces work together:

• Perception: Recognising the part, its position and its condition with cameras and sensors.
• Decision: Combining the image with the assigned task to decide what to do.
• Motion: Turning that decision into a safe, collision-free robot movement.

In short, the goal is a robot that can adapt to changing conditions the way a person does.

Story 1: LG tested its humanoid robot in a washing machine factory

According to The Herald Business on 10 November 2026, LG Electronics explained at the Nvidia AI Day event in Seoul how it trained and validated its CLOiD robot in a real washing machine factory.

The chosen task stands out: hanging 4-kilogram washing machine top panels on hangers. As Park Jeong-seop, head of LG's robotics intelligence team, put it, repeating this task over and over was causing musculoskeletal disorders and fatigue among workers.

The training process has three steps:

1. Three-dimensional motion data is extracted from videos of skilled workers.
2. This data is combined with examples gathered by teleoperation and with simulation data to train a vision-language-action model built on Nvidia's GR00T foundation model.
3. Trials run in a virtual environment that is nearly identical to the real factory, then the robot moves to the real floor. Issues found there feed back into training.

According to the report, LG raised its robot training data target from 100,000 hours to 500,000 hours.

The most instructive part is the challenges LG openly named. The robot must perform a task not just once or twice but continuously over long periods. As production plans change, parts and layouts change too. In Park's words, a factory floor may look standardised, but it holds many variables.

Story 2: ABB makes it easier to teach robots to see

According to Robotics & Automation News on 9 November 2026, ABB Robotics expanded its OmniCore EyeMotion product with a tool called AI Robot Trainer. The aim is to train, deploy and continuously improve AI-powered robot vision applications without needing experts who build models from scratch.

The key points in the report are:

• Real images captured on the shop floor are turned into usable training data.
• Image capture, model training, deployment and improvement come together in one workflow.
• A guided step-by-step flow and a chat assistant help users handle changes in product variants, lighting and layout.
• It is described as compatible with cameras and sensors from any brand.

According to ABB's own statement, when combined with automatic collision-free path planning the solution cuts commissioning time by up to 90 percent and cycle time by up to 50 percent. Keep in mind that these are vendor claims and that "up to" describes the best case.

Use cases range from simple pick and place to bin picking, quality inspection, sorting and assembly.

Story 3: NTN and Yazaki test visual inspection together

According to an NTN announcement on 5 November 2026, the company launched a joint initiative with Yazaki focused on AI and robotics. As a first step, NTN's robotic module i-WRIST was installed at Yazaki's innovation facility in Japan. There the two companies are testing i-WRIST together with AI image processing technology developed by Yazaki for visual inspection.

The lesson here is simple: even large companies do not put new technology straight onto the production line. They first test it with their own parts in a controlled setting, then judge whether it fits real production.

Lessons from the three stories

Read together, these developments paint the following picture for a factory manager:

• Heavy, repetitive work comes first. As in the LG case, tasks that tire workers and affect their health are the natural starting point for physical AI.
• Machine vision is the most mature area. Quality inspection, sorting and part picking reach the floor much sooner than humanoid robots.
• Data is the real asset. LG turning skilled workers' movements into data and ABB building training data from shop floor images point to the same truth: without data from your own floor, the model does not work.
• Test first, then scale. In all three cases, controlled validation comes first.
• Reliability over time is still hard. A demonstration is not the same as running without pause for a full shift.

Practical first steps for small manufacturers

Buying a humanoid robot is not a realistic goal for most small businesses today. But the thinking behind physical AI can be applied now, in smaller and reachable steps:

1. Pick one bottleneck. A visual quality check, a tiring handling task or a sorting job with frequent errors is a good candidate.
2. Start collecting images. Regularly saving photos of good and faulty products lays the foundation for any system you build later.
3. Record expert know-how. Document, in writing and with images, how your experienced staff spot a defect.
4. Set up a small pilot. Test at a single station with a measurable goal. Define the effect on error rate, time and workplace safety in advance.
5. Involve your team. Explain from the start whose workload the system will ease. A setup that leaves final decisions to human approval builds trust.

Frequently asked questions

How is physical AI different from classic automation? Classic automation runs on fixed rules. Physical AI aims to interpret images and adapt to changing conditions.

Will humanoid robots soon be common in factories? The LG case shows that trials on real floors have begun. Yet LG itself says that long, continuous operation and adapting to changing layouts are still major challenges.

Where should a small business start? With more mature applications such as vision-based quality inspection and sorting, through a small single-station pilot.

Conclusion

The November 2026 news shows AI moving from the screen to the factory floor. LG is testing its humanoid robot on a real line, ABB is making it easier to teach robots to see, and NTN and Yazaki are validating visual inspection in a controlled setting. The shared message is clear: the winners will be the businesses that collect data from their own floor and move forward with small, measurable pilots.

If you want to identify a task in your factory that machine vision or automation could ease and plan a small pilot, the AtomAxis team can help you with that first step.