Why this category is the honest benchmark
I spend most of my time on systems that produce text, and the hardest thing about evaluating them is that the grader is a person. A fluent wrong answer scores well with a reader who does not already know the answer. That is the whole reason calibration matters more than accuracy in that world.
Physical AI does not have that problem. The robot either picks up the part or it does not. The pallet is either stacked or on the floor. You cannot write a persuasive robot. Which is why, if you want an unsentimental read on how far the underlying technology has actually come, this is the category to watch rather than the one to ignore.
The constraint is data, and it does not scale the way text did
The last decade of AI progress ran on a specific piece of luck: humanity had already written down a very large amount of text and put it on the internet. Nobody collected that corpus for machine learning. It was free, and it was the input to everything that followed.
There is no equivalent for manipulation. No one has recorded a billion hours of hands picking up unfamiliar objects with the joint positions and force readings attached. So the data has to be manufactured, and the two available methods each carry a hard cost:
- Teleoperation. A human wears or drives the robot through the task. This produces high-quality data at exactly human speed and roughly human cost per hour. It works, and it does not get cheaper with scale in the way scraping does.
- Simulation. Generate the experience synthetically and transfer it to hardware. Vastly cheaper per hour, but what transfers is limited by how well the simulator models contact, friction, deformation and the long tail of the physical world. The residue is the sim-to-real gap, and closing it is the central research problem of the field.
This is the structural reason world models attract the research attention and the capital they do. They are the bet that a system can learn enough physics to generate its own training experience. If that works, the data constraint loosens. If it does not, progress stays coupled to how many hours of human demonstration someone is willing to fund.
What actually blocks deployment, in order
The public discussion treats capability as the gate. In the deployments I have seen described in any detail, it is usually third at best.
1. Cycle time
A robot that completes a task correctly but well below human speed does not replace a person; it becomes a second process running alongside one. The economics only close when the system approaches human rate or runs unattended for hours a human would not. Most current humanoid deployments are demonstrably below human rate on the single task they perform, and that number, not the demo footage, is the one to ask for.
2. Unit cost against the labour it displaces
The bill of materials for a general-purpose humanoid is dominated by actuators, sensors, and the hands. These are the components with the least commoditisation and the longest supply chains, and they are exactly the ones you cannot cheapen without losing the capability you built the machine for. Cost per unit needs to fall by close to an order of magnitude before the payback period looks like anything a CFO signs off, and that fall depends on volume that depends on the cost falling. Chicken-and-egg problems resolve, but slowly.
3. Task generality
A system trained on one task in one cell is an expensive single-purpose machine, and industrial automation already has cheaper single-purpose machines. The entire investment case for a humanoid form factor rests on generality: the same unit doing the next task without a re-engineering project. That is the capability question, and it is genuinely unresolved.
4. Safety standards that do not cover the case
ISO 10218-1:2025, published in January 2025, governs industrial robot safety, and the third edition folded the collaborative-operation requirements that previously lived in ISO/TS 15066 directly into the standard, alongside new functional-safety and cybersecurity clauses (A3's revision FAQ covers what changed). It assumes a robot in a defined workcell with characterised hazards. A mobile, general-purpose machine moving through an unstructured space occupied by people does not fit that frame, and there is no settled equivalent for it.
This one deserves more attention than it gets, because it is independent of capability. A system can work and still not be deployable, and no amount of model improvement resolves a standards gap. It resolves on committee timelines.
What this means if you run physical operations
The instinct is to evaluate vendors. I would do something less exciting first.
Every one of these systems is constrained by training data from your environment: your objects, your layout, your edge cases. That data is cheaper to start accumulating before you need it than to commission afterwards under time pressure. Instrumenting the processes you would eventually want automated is useful whether or not you ever buy a robot, because the same instrumentation tells you which processes are actually worth automating. Most organisations discover the answer is fewer than they assumed.
The second thing worth doing is separating the two questions people merge: can this technology do the task, and can this technology do the task better than the alternative. Physical AI is a subset of a more general decision about when a learned system beats a deterministic one, and the answer in a warehouse is often the same as the answer in a workflow: if you can enumerate the steps, do not buy something that infers them.
In this cluster
- World models — the research bet that would loosen the data constraint
- Embodied AI — the academic framing, and how it differs from the industry one
- Humanoid robot companies — who is building what, and what each has actually demonstrated
- Nuclear microreactors for AI data centers — Valar, Radiant, Aalo, Oklo and who is actually ahead
- AI infrastructure — the AI data center cluster: power, cooling, neoclouds and the largest sites
- Robots as a service — the commercial model that routes around the capital cost problem
- AI limitations — the general version of the argument on this page
Frequently asked questions
What is physical AI?
Physical AI is the application of machine learning to systems that act on the physical world rather than producing text, images or code: robots, autonomous vehicles, and machinery that perceives its surroundings and takes action in them. The term became common around 2024 as a way to distinguish this from generative AI. The distinction that matters is not the terminology but the feedback signal: a physical system is graded by the world rather than by a human reader, so its capability claims are far easier to check.
How is physical AI different from generative AI?
Three ways that change the engineering. Generative AI trains on abundant scraped data; physical AI has no equivalent corpus and must collect its own, largely through teleoperation, at human cost per hour. Generative AI failures are usually recoverable, since you regenerate. Physical failures damage things and injure people, so the acceptable error rate is orders of magnitude lower. And generative output is evaluated subjectively, while physical output is evaluated by whether the part actually moved.
What is the main limitation of physical AI right now?
Generalisation. Systems that perform well on a trained task in a known environment degrade sharply when the object, lighting, or layout changes, which is the problem the field calls the sim-to-real gap when the training happened in simulation. This is why deployed humanoid systems tend to run a single task in a fixed cell rather than the general-purpose assistant the marketing depicts. It is a data problem more than a model problem, which is why world models and simulation are where the serious research effort sits.
Are humanoid robots ready for commercial deployment?
For narrow, fixed tasks in controlled industrial settings, some are deployed today. For general-purpose work in unstructured human environments, no, and the gap is not primarily about intelligence. It is cycle time (current systems run well below human speed on the tasks they can do), capital cost per unit against the labour it replaces, and the absence of a settled safety standard for general-purpose robots operating around people.
What safety standards apply to robots working alongside people?
ISO 10218-1:2025, published in January 2025, is the current standard for industrial robot safety. Its third edition absorbed the collaborative-operation requirements that previously sat in the separate ISO/TS 15066 technical specification, and added functional-safety and cybersecurity clauses. It was written for robots in defined workcells with characterised hazards. A general-purpose mobile robot in an unstructured human environment does not fit that model cleanly, and no equivalent settled standard exists for it yet. That regulatory gap is an independent gate on deployment: a system can be technically capable and still not be deployable.
Should our organisation have a physical AI strategy?
Only if you own physical operations (manufacturing, logistics, warehousing, facilities) where labour is a material cost line. If so, the useful work today is not selecting a robot. It is instrumenting the processes you would eventually automate, because the constraint on every one of these systems is training data from your specific environment, and that data is cheaper to start collecting before you need it than after.