The definition, without the mysticism

A world model is a learned function. You give it the current state of an environment and an action you are considering. It returns its prediction of the resulting state. That is the whole idea, and the name it goes by in the older literature is model-based reinforcement learning. The current usage traces to Ha and Schmidhuber's 2018 paper World Models, which made the framing popular by training an agent entirely inside its own learned simulation of the environment.

The interesting property is what it enables: an agent holding such a model can evaluate an action without taking it. It can run a hundred candidate moves internally and act on the one with the best predicted outcome. In software that is a nice optimisation. When the actions involve a heavy machine near a person, it is the difference between a system you can test and one you cannot.

Why the idea came back

World models are not new, and their return in 2026 has less to do with a research breakthrough than with a constraint the field ran into.

Language models worked because the training data already existed. Nobody funded the writing of the internet; it was there, and scraping it cost approximately nothing per token. Robotics has no such corpus. There is no archive of a billion hours of hands manipulating unfamiliar objects with joint positions and force readings attached, and there is no way to scrape one, because the relevant signal was never recorded.

So the data must be produced. Teleoperation, a human driving the robot through the task, produces excellent data at exactly human speed and roughly human cost. It works and it does not get cheaper with scale. That is an unusual position for a field that has spent fifteen years assuming its inputs get cheaper.

A world model is the escape route. If a system can learn the dynamics well enough to generate its own experience, the marginal cost of a training hour collapses to compute, and robotics rejoins the scaling curve everything else has been riding. That is the bet. It is an economic bet dressed as a research programme, and the amount of capital behind it makes considerably more sense once you see it that way.

World model versus conventional simulator

Robotics has had simulators for decades, so it is fair to ask what is different.

A conventional simulator is written. Engineers encode the physics explicitly, and it is accurate exactly where the equations are good. It is weakest in the long tail (contact dynamics, deformable objects, friction on materials nobody characterised) which is unfortunately where most manipulation failures live. The residual error when you move a policy from that simulator onto hardware is the sim-to-real gap.

A world model is learned. It infers dynamics from observation, so it can in principle capture behaviour no one wrote an equation for. The trade is the one you always make when you replace a specified system with a learned one: the conventional simulator's errors are known and bounded, and the learned model's errors are neither. That is the same trade I described in agents versus automation, and it has the same shape here. You gain coverage of cases you could not enumerate, and you lose the ability to say confidently where the system will fail.

How to tell whether one is working

This is where I would push back on most of the coverage. The demonstrations that circulate are video: a model generates a convincing clip of an object being pushed, poured, dropped. The clips are genuinely impressive and they are weak evidence for the claim being made.

A video model trained on enormous quantities of footage has learned what plausible video looks like. Plausible video and correct physics overlap heavily, because most footage obeys physics, which is exactly why distinguishing them by eye does not work. The failure mode is familiar to anyone who has evaluated a language model: fluent output is not evidence of a correct model underneath. It is the same confusion, with better production values.

The test that means something is transfer. Train a policy entirely inside the world model, run it on real hardware, and measure the degradation. A small gap is a real result. Anything else (video fidelity, benchmark scores inside the model's own simulation, the length of a coherent generated sequence) is measuring the model against itself.

If you are evaluating a vendor or an investment in this space, that is the single question worth asking, and the answer is often either unavailable or narrower than the headline implies.

The same caution applies to the hardware end of the stack: a system can transfer perfectly and still be undeployable because it fails ISO 10218-1:2025, the 2025 industrial robot safety standard that now also carries the collaborative-operation requirements. Simulation quality and deployability are separate gates.

What follows if it works, and if it does not

If world models deliver transfer at scale, the data constraint on physical AI loosens and the deployment timeline compresses towards the remaining gates: unit cost, cycle time, and safety standards that have not been written. Those are real gates, but they are ordinary industrial problems with known resolution paths.

If they do not, robotics progress stays coupled to teleoperation hours, which means it advances at roughly the rate someone is willing to fund human demonstrators. That is a much slower curve, and it is the scenario the current investment levels are implicitly betting against.

Both outcomes are live. What I would avoid is treating the question as settled in either direction on the strength of a generated video.

Related

Frequently asked questions

What is a world model in AI?

A world model is a learned internal simulator of an environment. Given the current state and a proposed action, it predicts what happens next. That lets an agent evaluate options by imagining their consequences rather than trying them, which is both faster and considerably safer when the actions involve physical hardware. The term comes from a 2018 paper by Ha and Schmidhuber, though the underlying idea, model-based reinforcement learning, is older.

How is a world model different from a large language model?

An LLM predicts the next token in a sequence of text. A world model predicts the next state of an environment given an action. The distinction matters because language models learn how the world is described, and descriptions are systematically incomplete. Nobody writes down that an unsupported object falls, so the relationship between action and consequence is underrepresented in text. A world model is trained on that relationship directly.

Why are world models important for robotics?

Because robot training data is expensive in a way text never was. There is no scraped corpus of manipulation, so data is collected largely through teleoperation at roughly human cost per hour. A sufficiently accurate world model breaks that dependency: the robot generates its own experience, and the cost of a training hour falls to the cost of compute. Every serious robotics programme is pursuing some version of this, because the alternative is progress bounded by how much human demonstration time someone will fund.

Do world models actually work yet?

They work in the narrow sense and are unproven in the broad one. Learned simulators produce useful training signal in constrained, well-characterised settings, and video-generation models have become strikingly good at producing plausible sequences. What is not established is that this constitutes learned physics rather than learned visual plausibility. The test is transfer: does a policy trained inside the model work on real hardware, and how far does it degrade. Video quality is not evidence either way.

What is the difference between a world model and a simulator?

A conventional simulator is written by engineers who encode physics equations explicitly. It is accurate where the equations are good and wrong in the long tail: contact, deformation, friction on unfamiliar materials. A world model learns the dynamics from observation instead, so it can in principle capture behaviour nobody wrote an equation for. The trade is that its errors are unpredictable rather than characterised, which is a meaningful difference when the output controls a machine near people.

Are world models the same as physical AI?

No. Physical AI is the broad category of AI systems acting on the physical world. A world model is one technique inside it, and the one most likely to determine the pace of the rest, because it targets the data constraint rather than any particular capability.