Physical AI Can Act. It Still Can’t Check Its Own Work.

For ten months, a humanoid robot clocked in at BMW’s Spartanburg, South Carolina plant, ten-hour shifts, Monday through Friday. Figure AI’s Figure 02 moved more than 90,000 components and supported the production of more than 30,000 BMW X3s, BMW Group confirmed in February 2026. The assignment was narrow and unforgiving: pick sheet-metal parts and load them into a welding fixture, each placement inside a five-millimeter tolerance, in roughly two seconds, with a target of better than 99% successful placements per shift.

• A robot can return to the same commanded position all shift and still produce an out-of-tolerance assembly.
• Repeatability of 0.1 mm or better is common; absolute accuracy typically runs between 0.5 and 1.5 mm.
• Many industrial customers expect 99.99% reliability, and the cost of poor quality runs 10% to 15% of operations.

Even at that target, placement success and quality conformance are two different measurements. The five millimeters is a loading tolerance, not a product tolerance; the welding itself was done by fixed robots downstream. A humanoid can execute every commanded motion correctly and still hand off an assembly that ends up outside the engineering tolerance. The question that matters on the line is not whether the robot finished. It is whether the part is in specification, and whether anything independent of the robot established that.

The robots of science fiction were always arriving “someday.” They are now clocking hours. In the same February announcement, BMW Group said it would run its first European humanoid pilot at Leipzig, supporting high-voltage battery assembly and material delivery, and, more telling for anyone who works in quality, that it was standing up a Center of Competence for Physical AI in Production.

That phrase is worth pausing on. AI has left the screen. Machines now perceive, reason, move and act on physical material. Which raises a more immediate question than whether robots will replace people.

Who verifies that the work was done correctly?

A robot can complete a movement and still be wrong. It can set a component two millimeters off datum, finish an assembly outside tolerance, or execute a sequence that looks clean on the HMI and should fail inspection. “The robot did it” is not a measurement.

What Physical AI Actually Means

Physical AI describes autonomous systems that perceive, reason and act in the physical world using an AI world model, a learned internal representation of how objects, surfaces and forces behave. NVIDIA, which has built much of the compute layer underneath this shift, framed its CES 2026 announcements around the claim that “the ChatGPT moment for physical AI is here, when machines begin to understand, reason and act in the real world.”

The difference from software AI is consequential. A language model that errs produces a wrong sentence. A physical AI system that errs damages a fixture, stops a line, or produces several hundred parts that all have to be contained.

That changes what the system has to do. An industrial physical AI loop has one more step than the version most people describe.

Cameras and sensors read the environment. The model interprets it. The robot acts. Then comes the step manufacturing actually cares about: verification. Did the physical result match the engineering intent, and if it did not, what changes on the next cycle?

Why Acting Is Not Enough

Picture a humanoid positioning a bracket. It identifies the part, reaches it, moves it into place and completes its motion program, probably triggering the next operation. From the robot’s side of the transaction, the task is closed.

From a quality engineer’s side, the questions are just opening. Is the part located and oriented correctly? Are the critical features within tolerance? Does the result match the CAD definition?

The gap is measurable, and it is well characterized. Industrial robots are very good at repeatability; many achieve 0.1 mm or better. Absolute accuracy, meaning how close the tool center point actually lands relative to the commanded position, is another matter: typical values run between 0.5 and 1.5 mm, according to Martin Finkbeiner of the Fraunhofer Institute for Manufacturing Engineering and Automation. Both properties are defined and tested under ISO 9283, and confusing one for the other is a common way to over-trust a cell.

Repeatability is also only one contributor to the finished result. Part variation, fixture condition, end-effector compliance, tooling wear, datum stack-up, thermal drift and calibration all show up in the measurement. A robot can return to the same commanded position all shift and still produce an out-of-tolerance assembly because something upstream moved.

The industrial bar makes that margin unforgiving. Writing in IEEE Spectrum in September 2025, Evan Ackerman reported that many industrial customers expect a couple more nines than 99%, 99.99% reliability, because line-stopping downtime can cost tens of thousands of dollars per minute. Melonee Wise, who had served as chief product officer at Agility Robotics until that month, said the company had demonstrated that level in some specific applications, but not in general-purpose ones. Her assessment of the broader field was blunt: “the reality of the situation is that currently AI is not robust enough to meet the requirements of the market.”

This is why process qualification does not ask whether the robot reached its programmed position. It asks whether the critical characteristics on the part stay within tolerance at production cycle time, measured by a system capable enough to make that call reliably.

Putting the Verification Layer on a Moving Robot

A recent trial by SkillReal with a leading humanoid robot manufacturer was built around exactly that gap. Rather than adding a dedicated measurement station alongside the robot, the trial fed the humanoid’s own camera stream into SkillReal’s 3D-AI inspection platform.

That is possible because of how the platform is constructed. Proprietary sub-pixel calibration turns standard industrial cameras into 3D sensors, and pre-trained models, built on more than 35,000 real manufacturing variations, with no part-specific training, locate the part in the scene. The part is then aligned to its CAD definition and measured against GD&T requirements, without jigs or fixtures. SkillReal’s platform measures features against their CAD definition inside the production cycle. Precision down to 0.05 mm. Performance depends on the application and operating conditions. It runs on standard PCs and edge GPUs.

The humanoid’s onboard cameras were designed for robotic perception, not dimensional metrology. The trial calibrated that camera system, processed its image stream to raise its quality, aligned the CAD definition to the observed part and measured selected features against their tolerances. It did this without adding a separate metrology sensor. Moving a camera closer puts more pixels across a feature and improves measurement sensitivity, and a humanoid, unlike a fixed camera, can do exactly that. The machine that acts should not be its only verifier. The layer riding on its camera can be the second one.

So, the workflow inverts. The robot observes. The part’s true position is calculated. The selected features are checked against tolerance. The system decides, records the result, and feeds the measurement back, at which point the robot acts again, sending the part to the repair bin or on to the conveyor. That same feedback can be used to calculate offsets that compensate for a detected deviation on the next cycle, rather than waiting for an end-of-line gate to find it.

Closing the Loop

Traditional automation starts with assumptions. The part should be here. The robot should move there. This is the result it should produce.

Physical AI replaces those assumptions with questions, then measures the answers. A robot completing a programmed movement is automation. A system that measures the part, judges conformance, feeds that judgment back and selects the next action is moving toward autonomy.

The flexibility of humanoids is impressive, but production still needs a verification layer. The same low-weight 3D-AI inspection technology we are already using on demanding automotive production lines can move with a robot, measure a part against its CAD definition, and turn the result into an immediate quality decision and action.
— Shai Newman, CEO, SkillReal

AI supplies awareness and reasoning. Robotics supplies movement. Manufacturing still needs verification, which is why dimensional inspection belongs between an AI decision and a result anyone is willing to trust.

A verification layer does not get a pass because it uses AI. For quality teams the questions are familiar. What are the repeatability and reproducibility? What is the measurement uncertainty relative to the tolerance? How does the system correlate with reference metrology, and how are software and model changes controlled? Physical AI does not lower the metrology bar; it moves that bar into a more flexible production control loop. Those are the questions a plant should put to any verification layer, this one included.

SkillReal describes its longer-term role as the verification layer for physical AI, connecting physical geometry to engineering intent across fixed cameras, robots and, increasingly, mobile platforms. The framing has drawn infrastructure-level attention: the platform is accelerated on NVIDIA’s AI stack, and the company is a Siemens partner for digital-twin manufacturing.

The underlying idea is not complicated. Do not only give machines the ability to act. Give them a reliable way to verify the result.

The Same Problem, Without the Humanoid

The verification problem is not unique to humanoids, and neither is its cost. ASQ’s technical report on cost of quality (TR2:2024) puts the cost of poor quality at 10% to 15% of operations even at a thriving company, repair, rework, scrap, warranty and write-offs included. Yet many plants still lean on offline sampling: a limited number of parts measured per shift, usually well after a process has started to drift.

The value of inspecting every part goes well beyond pass/fail sorting. When dimensional and feature data are collected continuously, quality stops being a gate and becomes a process signal. A plant can watch a critical characteristic begin to move, correlate the trend to tooling, fixtures or other process conditions, and intervene before the shift produces a containable quantity of nonconforming parts. That is a different activity from discovering the problem hours later in an offline sample.

The same platform is already in production at OEMs and Tier-1 suppliers in Japan and the United States, measuring hundreds of features per part within production cycle time, on every part rather than a sample, more than 80 million features a month across the installed base. Because the architecture is software-defined and built around industrial cameras and CAD-based inspection, it can often be retrofitted into an existing line without adding a dedicated metrology cell or a part-specific hard gauge.

Integration work does not disappear. Cameras and lighting have to be positioned correctly, controls communication established, compute integrated, and the measurement system validated against the requirements of the application, a gauge R&R study is still a gauge R&R study. What changes is how much station-specific metrology hardware, custom fixturing and engineering effort each installation demands.

That distinction is what decides whether a pilot becomes a program. A technology can perform well in a demonstration and still be difficult to scale if every station turns into its own engineering project. A software-defined inspection architecture can be replicated across stations and lines, which lets a plant start with one high-value application, validate performance and the business case there, and expand from a known quantity. According to SkillReal, several existing deployments have followed that path into larger multi-system installations.

Manufacturers often underestimate deployment friction. A technology can be accurate in a demo and still fail economically if every station requires custom sensors, fixtures, programming and weeks of integration. The breakthrough is inspection that is CAD-based, metrology-grade, practical to integrate with standard industrial cameras, and scalable enough to deploy across a plant.
— Shai Newman, CEO, SkillReal

What to Ask of Any Physical AI System

The principle applies well beyond humanoids, to robotic arms, autonomous inspection and flexible cells. Three capabilities are worth putting on the evaluation sheet now.

• Real-world adaptability. The machine has to respond to the environment that exists, not the one that was programmed.
• Metrology-grade verification. The system has to distinguish between completing an action and producing a result inside each feature’s tolerance, and prove that distinction with a capable measurement system, not an inference.
• Low-touch, scalable deployment. The inspection layer has to connect physical reality to CAD without turning every station into a custom metrology project.

Together they describe a plant where inspection is no longer a downstream gate but a continuous feedback layer: observing, measuring, deciding, correcting.

The Robots Are Here. Trust Is the Next Problem.

Humanoid robots attract attention because they look like the future we were promised, or warned about. The form factor is the least interesting part. The real shift starts when machines move from executing predefined motions to making decisions in environments that are variable, imperfect and physical. That is where risk accumulates: the more authority we hand over, the more it matters to know what was actually done.

The humanoid trial is one early view of that transition, robotics supplying capability, AI reading the environment, inspection tying the outcome back to engineering intent. The more useful point is that the verification layer does not have to wait for humanoids. It is on production lines now, solving a problem most plants already have.

Manufacturers do not have to choose between today’s quality bottleneck and tomorrow’s autonomy. The CAD-driven inspection layer that clears the first is the same feedback layer physical AI will need as machines take on more responsibility.

The question is no longer whether machines can do the work. It is whether manufacturers can verify that work accurately, continuously and economically enough to trust it.