Arrivals0
Groundings0
Position error0.00 nm
Since last fix0 min
A field guide to physical AI and world models

A world model is dead reckoning for machines.

Before satellites, a navigator worked out a ship's position from heading, speed and the hours since the last landmark. Robots still do it. A world model extends the habit from position to everything else in the scene.

Scroll to follow the vessel
On the chart · Fan of tracks

Rollouts

At every fix the vessel imagines forty futures. Each pink line is one of them, played forward inside the model without the vessel moving at all. Planning is choosing among them, and the solid line is the choice.

This view: fix every 3 min, imagines 6 nm ahead

On the chart · Current

Unmodelled dynamics

The arrows are a current the model has never heard of. In a robot it is friction, cable slack, a box heavier than yesterday's. The plan assumes still water, so the vessel drifts off the line it believes it is sailing.

This view: current raised to 3 knots

On the chart · Circle of uncertainty

Compounding error

The ring marks where the model thinks the vessel is, and how wrong that could be. It widens with every minute since the last fix, because each predicted step inherits the error of the one before.

This view: fix every 14 min

On the chart · Fix

Observation

A fix is a look at the world: a camera frame, a force reading. The ring collapses, the model replans from the truth, and the drift starts again. The more often a robot can afford to look, the worse its model is allowed to be.

This view: fix every 4 min in a 2.5 knot current

On the chart · Grounding

Failure in the world

This crew checks its position every thirty minutes in a four-knot current. The plan stays sound inside the model while the vessel runs onto the rocks. Stay here a moment and watch the groundings counter.

This view: fix every 30 min, 4 knot current

Your turn

Take the helm.

Click the water to set a destination. Drag to turn the chart.

Fixes come often and the model sees far ahead. The current still pushes, and each fix absorbs it.

  • Track sailed
  • Futures imagined
  • Future chosen
  • Believed position
The arithmetic of a long task

A demo is a short rollout. A shift is a long one.

Set how often one step works and how many steps run without a check. Each row below is one simulated run, and it stops at its first failure.

37%of runs finish clean

Each mark is one step. Pink marks the first failure.
Why now

Language models learned from text people had already written. Physical AI has no such archive. The record of what happens when a gripper closes on a wet glass has to be collected by robots, or imagined by a model good enough to stand in for them. World models moved to the centre of the field because they give a machine a cheap place to practise.

The stack

A physical AI system has to see, predict, act, practise and prove it.

Five layers, from the sensors up. Select a layer to read what it does and what to ask about it.

Layer 2 of 5

Predict

For your next vendor meeting

Five questions to put to anyone selling physical AI.

  1. Which senses does the policy use at run time, and how often does it read them?
  2. Does the system predict consequences before it moves, and how many steps ahead?
  3. What share of the training data came from robots like the one on offer?
  4. How much of the training was real, and how was the transfer from simulation measured?
  5. What was the success rate over the last thousand unattended cycles, and who counted?
Deployment

The charted water is narrower than the demos suggest.

For a decade the pattern has held: deployment arrives first where the environment is bounded and a mistake can be retried.

Charted

In commercial service

  • Driverless ride-hailingPaying passengers in more than a dozen US cities, inside mapped service areas with remote support on call.
  • Warehouse picking and sortingArms pulling mixed items from bins across full shifts. The task is bounded and a miss gets retried.
  • Visual inspectionLearned models checking welds, labels and surfaces. Nothing moves on the model's say-so, which keeps errors cheap.
475
Coastal

Paid pilots on narrow tasks

  • Humanoids on factory floorsMoving totes and loading fixtures in a few plants and warehouses, supervised, and slower than the people beside them.
  • Driverless freightTrucks on fixed highway routes between depots, in fair weather.
  • Mobile arms in hospitals and labsFetching and delivering along corridors the robot already knows.
38715463
Open water

Still research

  • General home robotsEvery kitchen is a new environment. Many of the impressive demonstrations have a person at the controls.
  • Soft and wet thingsCloth, cables, food. Touch sensing and simulation both fall short here.
  • Hours without supervisionError compounds over a long task, as the arithmetic above shows.

The placement is a judgment call, current to October 2026.

The field

Thirty-six years from an idea to an $8.2 billion acquisition.

Each mark is a system or a moment worth knowing. Select one to read about it. The 2025 column is where the large labs committed.

Readiness

Read the barometer before you commit to a pilot.

Pick one physical task and answer for that task alone. The answers loaded here are an example for a parcel-sorting line. Replace them with yours.

Passage plan

Sail a first pilot in legs, and leave each one on evidence.

From harbour to open water. Write each exit condition down before the leg starts.

  1. Leg 1

    Record

    Instrument the station. Capture video and machine actions together, with a pass or fail on every cycle.

    Leave whenyou hold a few hundred labelled cycles that cover the ordinary variation.

  2. Leg 2

    Rehearse

    Adapt a policy to the task and run it on a test rig or a simulated copy of the station, away from production.

    Leave whenmore data stops improving the success rate on the rig.

  3. Leg 3

    Shadow

    The model watches live production and proposes actions that nobody executes. Compare each proposal with what the operator did.

    Leave whenits misses are few and every one of them has an explanation.

  4. Leg 4

    Supervise

    The machine acts while a person stands by with a stop button. Count every intervention.

    Leave wheninterventions per shift fall below the number you wrote down before starting.

  5. Leg 5

    Release

    The machine works alone inside fixed limits, with an automatic check at the point of work.

    Thenwiden one variable at a time and repeat the count.

Terms

If a term cannot be unpacked, it is decoration.

Nine words that come up in every conversation about physical AI, one layer down.

World model
A learned function that takes the current state of an environment and an action, and returns the next state. A system that cannot be asked "what happens if I do this?" is a video generator.
Vision-language-action model
A policy that reads camera frames and a written instruction and writes motor commands. The language part usually comes from a pretrained model, and the action part is trained on robot demonstrations.
Rollout
A sequence of predicted steps, each fed from the one before. A planner scores many rollouts and executes the opening moves of the best one. The fan on the chart is forty of them.
Model predictive control
Plan over a horizon, act on the first part of the plan, observe, plan again. The vessel on the chart does exactly this, and so do most robots that use a world model at run time.
Latent space
The compressed internal description a model keeps of a scene. Predicting there is cheaper than predicting pixels, at the cost of being harder for a person to inspect.
Sim-to-real gap
The drop in performance when a policy trained in simulation meets real surfaces, lighting and wear. Teams narrow it by randomising the simulation and by mixing in real data.
Teleoperation
A person drives the robot through a task while every frame and joint command is recorded. Most action-model training data is still made this way, one demonstration at a time.
Cross-embodiment
Training one policy on data from many robot bodies so that skills learned on an arm with two fingers carry over to a humanoid hand.
Digital twin
A simulated copy of a specific site, kept in step with the real one. A generic simulation teaches a skill, and a twin lets you rehearse it on your own floor plan.
Reading

Sources worth an evening.

Primary sources, from the paper that named the idea to last week's acquisition.