Physical AI and the Unpredictable Factory

ReadIndustry InsightsStats & TrendsBest PracticesInfographicCool Animations
Written by Jeff Winter

AI is leaving the chat, and someone has handed it a forklift, which makes for a much more interesting conversation once you consider everything that can go wrong between “pick up that pallet” and actually picking up that pallet. The wrapping is loose, the load is uneven, somebody has parked in the aisle, and the machine is now expected to deal with a situation that looked considerably tidier in the presentation (the pallet, regrettably, has not read the project charter).

That is the part of Physical AI that interests me. I spend a lot of time thinking about manufacturing, AI, and digital infrastructure, and I keep coming back to the work required to make technology useful when conditions change. NVIDIA’s definition covers machines that perceive, understand, and act in the physical world. My interest is in how much variation they can handle, what people still have to do for them, and whether the operation improves as a result.

The term has a history, even if the attention feels new

In November 2020, Aslan Miriyev and Mirko Kovač published “Skills for physical artificial intelligence” in Nature Machine Intelligence, describing the joint development of robot bodies, materials, sensing, actuation, and control. Their argument put physical design inside the discussion about intelligence, with a much broader scope than adding a language model to an existing robot.

NVIDIA deserves credit for giving the term a much larger audience. At COMPUTEX in June 2024, Jensen Huang called physical AI “the next wave of AI”; the company followed with Cosmos in January 2025, bringing world foundation models and development tools into the story. March brought Google’s Gemini Robotics and NVIDIA’s GR00T N1, while NVIDIA’s CES 2026 announcements expanded its Physical AI models and infrastructure. Taken together, those releases help explain why I see this as both a research story and a commercialization story. For a manufacturer, I think the useful question is how much of that development work can be reused across applications, and how much still has to be engineered for every new task. That distinction belongs in the business case before anyone starts counting robots.

The Google Trends export accessed October 26, 2026 adds some perspective. Worldwide search interest for “Physical AI” rose from an index of 3 in January 2025 to 32 that September, reached 100 in June 2026, and stood at 49 in September. I have excluded October 2026 because the month is incomplete. Google’s methodology matters here: these are relative interest scores, with 100 marking the peak, rather than counts of searches or deployments. The announcements are useful landmarks; this chart cannot tell us which one caused somebody to start searching.

For me, the interesting question is what sits underneath that attention. The five parts of the infographic offer a useful way to unpack it, so let’s follow one carton through the process.

Perceive: “I see a box” leaves quite a lot unresolved

When a carton approaches a robot, identifying it as a carton barely gets the job started. Its position, angle, surrounding obstacles, and condition all affect what a useful next action might be, which is why the infographic begins with perception rather than movement.

NVIDIA describes Physical AI systems using multiple inputs, including images, video, and sensor data. In our example, a camera might locate the carton, depth sensing might help estimate its shape, and force feedback might reveal what happens when the gripper makes contact. Those are complementary clues about the same physical situation, with the particular combination depending on the application.

I care less about the colorful outline around the object on a demonstration screen than whether the machine notices something that changes its decision. A damaged corner should matter if it affects the grip; an obstructed view should matter if the system no longer has enough information to proceed confidently. Try the actual packaging under different lighting, partly hidden behind another carton, and at the awkward angle the demonstration somehow avoided. That is where I would start learning what the perception system can be trusted to support. I would record results separately for each troublesome condition, including how often the system correctly asks for help. Otherwise, hundreds of easy picks can make the average look excellent while concealing a persistent problem with one supplier’s glossy packaging. Keep examples of those failures alongside the score, so the next discussion starts with what actually confused the machine.

Predict: Think through the move before gravity gets involved

Now we have an estimate of the scene, but several actions may still be possible, and their consequences could be very different. Gripping near an edge might leave the carton unstable, while the most direct route could bring it too close to another object; choosing well requires some way to anticipate what follows from the movement.

World models are one approach. NVIDIA’s Cosmos announcement describes generating possible future scenarios and supporting model evaluation, giving developers ways to explore outcomes before physical execution. Other approaches organize the reasoning differently: Gemini Robotics-ER combines spatial understanding with the ability to work through existing controllers. The infographic’s five functions are a way to examine the system, not a claim that every robot contains five neatly separated software boxes.

This is also where I want the manufacturing team involved. Ask which assumptions the model makes about weight, stiffness, friction, and movement, then bring the supplier variation and damaged packaging into testing. A beautiful simulation is still making assumptions about the factory, and I would much rather discover the wrong ones before production depends on them (gravity has shown very little interest in our implementation schedule). Ask the team to show which assumption would change the chosen movement; if nobody can explain that, I would want more evidence before trusting the prediction.

Act and adapt: The demonstration should keep running after something slips

Suppose the gripper lifts the carton successfully, then the carton shifts halfway through the movement. I would happily watch that part of a demonstration for longer, because recognizing the change, correcting within tested limits, and deciding when to request help tell me something about the work people will inherit after installation.

Google DeepMind describes Gemini Robotics as monitoring its surroundings and adjusting its behavior when conditions change, including replanning when an object slips or is moved. Vision-language-action models connect visual observations and instructions to physical actions, allowing the machine to respond as the task unfolds. The progress is worth taking seriously without turning every benchmark into a production promise. In its 2023 RT-2 research, DeepMind reported performance on previously unseen scenarios improving from RT-1’s 32% to 62%. That is a substantial improvement within the evaluation, although nobody should mistake it for a factory acceptance test.

My questions would quickly move to completed tasks, damaged items, intervention frequency, and recovery time. If a person has to rescue the machine repeatedly, include that work in the result, along with restarting, diagnosing, and recovering whatever it dropped. Consider a hypothetical operation attempting 10,000 picks a day: 99% completion without human help still leaves 100 exceptions. At three minutes of staff time per exception, that is five hours of work, before counting any production delays. I would also ask when those interventions occur, because several machines asking for the same person at once creates a different staffing problem from evenly spaced requests. Count retries, too, and separate eventual completion from completion within the required cycle time. A robot that eventually gets there may still leave the next process waiting (a distinction the delivery schedule will be delighted to explain).

The value of intelligence is measured by what changes because of it, and a faster pick has limited value if it creates another job for somebody downstream.

Learn and transfer: Experience has to travel beyond the training setup

One reason I find this field interesting is how much effort goes into giving machines experience that people accumulate almost without noticing. The DROID project assembled 76,000 demonstration trajectories, representing 350 hours of interaction across 564 scenes, with fifty collectors contributing over twelve months. That is a considerable amount of work behind the apparently simple idea of teaching a robot to handle things.

Simulation offers another source of experience, allowing developers to vary conditions and evaluate behavior before deployment. NVIDIA’s three-computer framework connects training, simulation and validation, and execution on physical machines, with operational data feeding subsequent development. For our carton, the exercise could include different sizes, orientations, surfaces, and lighting, followed by testing on the equipment and materials the factory actually uses. In its March 2025 GR00T announcement, NVIDIA reported generating 780,000 synthetic trajectories in 11 hours, which it described as equivalent to 6,500 hours of human demonstration data. That helps explain the excitement around simulation, although the comparison concerns the quantity of generated experience. I would still want to see which physical trials were used to check that the simulated experience taught the right behavior.

DeepMind reported in March 2025 that Gemini Robotics more than doubled average performance on its generalization benchmark compared with other leading vision-language-action models. Generalization is the attraction here: useful behavior should extend beyond the exact situations represented in training. How far it extends in your application still needs to be demonstrated. Reserve some packaging variants and workcell arrangements for evaluation only, then check whether performance holds up when those unfamiliar conditions appear. After an update, rerun the tasks that already worked. An improvement on the awkward new carton is less impressive if the machine has quietly become worse at handling the ordinary ones.

There is also a distinction worth preserving between a machine adjusting its next movement and its trained model being updated, as illustrated by DeepMind’s separate discussion of on-device execution and fine-tuning. I want model changes evaluated before they return to production, with operators helping choose the difficult cases. Ask them which exceptions the standard procedure leaves out; their answers are likely to make the test considerably less comfortable and considerably more useful.

Digital infrastructure: The part I have a professional interest in

My work at Belden makes me particularly interested in the center of this infographic, because networks, compute, data, and security have to support everything happening around them. You can spend a great deal of time discussing what the model understands while leaving rather important questions about how the system operates unanswered.

For example, DeepMind designed Gemini Robotics On-Device to run locally, supporting latency-sensitive applications and environments with intermittent or no network connectivity. That is an architectural choice with consequences for how and where the machine can operate. It also gives us something specific to discuss beyond the increasingly unhelpful instruction to “put AI in it.”

I want to know which connections are essential, where time-sensitive processing happens, who can change the model, and whether maintenance can reconstruct what the machine observed before a failure. Someone should also be able to restore a previously validated configuration without assembling the original development team (apparently equipment is allowed to break outside the project meeting). A useful acceptance test would follow one decision from the camera timestamp through the model output to the resulting machine action. Can the team establish how old the observation was when the robot used it, and recover the model version and relevant sensor readings afterward? I would also want a controlled test of lost connectivity, with the expected behavior agreed in advance. Which tasks can continue locally, which must pause, and how will interrupted work be reconciled when the connection returns?

Coordination matters as well. In its 2025 DeepFleet announcement, Amazon reported deploying its millionth robot across more than 300 facilities and anticipated a 10% improvement in robot travel time through fleet coordination. That was its expectation at launch, but the target is interesting: better movement across the operation. A faster individual machine accomplishes little if the next station cannot accept its output, so any plan for additional capacity should include what the rest of the process will do with it.

What I would actually want to change

Start with work where variation carries a measurable cost, establish the current performance, and let the people who understand the exceptions challenge the proposed solution. Include engineering and ongoing support in the comparison, and keep ordinary automation in consideration when it already solves the problem reliably. Before choosing a pilot, spend a week recording the exceptions in one candidate process and the minutes required to resolve each. Include setup changes and engineering visits, because adaptation already has a cost even when nobody calls it that. Use that record to choose a task with enough recurring difficulty to matter commercially and a result the operating team can measure.

The deeper opportunity with Physical AI is reducing how much of the world we must make predictable before machines can be useful. If a change in packaging, product mix, or operating conditions requires less reprogramming and less human intervention, the economics of previously difficult work could begin to change. That is the possibility I find compelling: equipment that remains useful as the work changes around it, giving people more room to deal with the problems that still need their judgment.


References:

Jeff Winter
Next

Congratulations! Your Plant Has Been Selected for Digital Transformation.