What Is a World Model? How AI Learns to Imagine What Happens Next

Why it matters now

On 27 October 2026 Google DeepMind announced Genie 4, which it calls “our new world model”. You steer through a scene with the keyboard, type prompts to change the weather or add people, and the model generates what you see and hear in real time. DeepMind lists “training environments for robots and agents” among its uses. We covered the launch here.

The same week, the start-up Dream Labs opened a paid API for Dream-1, a robot brain that predicts video and robot movements together, and the Chinese robot maker AGIBOT previewed GO-3, a model it says helps robots “anticipate the results of their actions” (Dream Labs brief, AGIBOT brief).

So the phrase “world model” is turning up in video games, chatbots and robotics at once. This explainer covers what one is, how it works and where it still goes wrong.

How it works, step by step

Step 1: Learn by watching

A world model starts the way a baby does: by watching. It is trained on huge amounts of video, and sometimes on recordings of actions such as button presses or robot arm movements. Meta’s V-JEPA 2 world model, for example, was pre-trained on “over 1 million hours of internet video”.

Nobody writes rules such as “things fall down” into it. The model works the rules out for itself, because getting the next moment right is much easier if you have picked up how gravity, collisions and shadows behave.

Step 2: Squeeze the scene into a summary

A single video frame holds millions of numbers. Predicting every one of them is slow and wasteful, so most world models first compress the scene into a much smaller internal summary: roughly “a red cup, near the edge of a table, a hand moving towards it”. Researchers call this summary a latent state. It is the model’s private notes about what matters in the scene.

Step 3: Predict the next moment

Now comes the core trick. Give the model its summary of the scene plus an action, and it predicts the summary of the next moment. Then it does it again, and again, rolling the future forward one step at a time.

An everyday analogy: think of a chess player who thinks three moves ahead. They do not move the real pieces. They picture the board after each possible move, and after the opponent’s likely reply. A world model does the same for the physical world, except that its board is a kitchen, a street or a game level.

Step 4: Turn the prediction back into pictures (optional)

Some world models, like Genie 4, turn each predicted moment back into video and sound so a person can see and hear it. That is what makes them feel like an interactive game. Some use techniques related to diffusion models, the same family behind AI image generators, to paint each frame.

Others never draw a picture at all. A robot does not need a pretty video; it only needs to know whether its plan ends with the cup in the sink or on the floor.

Step 5: Practise and plan inside the imagination

This is where world models become useful. An agent can try thousands of actions inside the model, see which predicted futures go well, and only then act for real. In 2018 the researchers David Ha and Jürgen Schmidhuber wrote that they could “train our agent entirely inside of its own hallucinated dream” generated by its world model.

The idea has grown up since. The DreamerV3 system, published in Nature in 2025, learns a world model of a game and practises inside it; its authors describe it as “the first algorithm to collect diamonds in Minecraft from scratch without human data or curricula”.

A worked example: a robot clearing a table

Imagine a two-armed robot told “put the cup in the sink”.

1. Its camera sees the scene. The world model compresses it into a summary: cup, table edge, sink, its own two grippers.
2. It imagines a few short plans. Plan A: reach straight for the cup. Plan B: slide the cup away from the edge first. Plan C: use the other arm.
3. It rolls each plan forward. In the imagined future of Plan A, the gripper clips the cup and it tips off the table. Plan B ends with the cup held firmly. Plan C collides with a plate.
4. It picks Plan B and carries out only the first part, then looks again and repeats. Short steps matter, because each new camera image corrects the model’s mistakes before they pile up.

This is close to how Meta describes its V-JEPA 2 robot version: after training on “less than 62 hours of unlabeled robot videos”, the team deployed it on robot arms in two labs it had never seen, with no task-specific training, and had it plan by imagining the result of candidate actions.

What people get wrong

“It is a physics engine.” A video game engine follows physics rules that programmers wrote down. A world model has only learned patterns from data, so it can be confidently wrong. DeepMind says Genie 4 keeps scenes coherent “over short sessions”, and that visual details and object permanence “can still shift or drift over time”. In plain words: look away and a door might move.

“Realistic video means it understands the world.” Looking right is not the same as being right. A model can draw a convincing glass of water and still get wrong how the water spills.

“It is just a chatbot that makes videos.” A large language model predicts the next word. A world model predicts the next state of a scene, usually given an action. The two are increasingly combined: in Genie 4, a Gemini model acts as a guide that adapts prompts to the scene.

“Robots trained in imagination are ready for the real world.” Lessons learned in a simulation do not always survive contact with real floors, lighting and friction. Closing that gap is called sim-to-real transfer, and it is still hard.

Where it is heading

• More people will get to try them. Google says Genie 4 is a research preview for trusted testers, coming to Google AI Ultra subscribers and early access partners in the Gemini API and Google AI Studio, without giving dates.
• Robot brains are being rented, not just built in. Dream Labs prices Dream-1 from $3 an hour and calls it a research beta that “has not been validated for safety-critical use”.
• World models and action models are merging. Dream-1 is what its maker calls a world action model, predicting video and movements together. Our world-action models explainer covers that idea in 2 minutes.
• The big open problem is memory. Longer, consistent sessions are what every lab is chasing, because a world that forgets itself is no use for training a robot over a full task.

Key terms

• World model
• Sim-to-real
• Vision-language-action model
• Embodied AI
• Diffusion model
• Multimodal
• Watermarking
• Latent state

Sources

• Google DeepMind, “Genie 4: A new generation of world simulation”, 27 October 2026. deepmind.google
• Ha and Schmidhuber, “World Models”, arXiv:1803.10122, 2018. arxiv.org
• Hafner et al., “Mastering Diverse Control Tasks through World Models”, Nature, 2025 (project page). danijar.com
• Assran et al., “V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning”, arXiv:2506.09985, June 2025. arxiv.org
• Dream Labs, Dream-1 research page, as reported in our brief of 2 November 2026.
• AGIBOT, GO-3 Preview announcement, 29 October 2026, as reported in our brief.