Danijar Hafner: learning to act through world models

Learning what an action might do

For an agent looking through a camera, recognising an object is only the beginning. To act, it must deal with change: a movement alters what it sees, an object may move out of reach, and a promising choice can have consequences much later. Danijar Hafner’s research asks how machines can learn representations that help them anticipate those consequences and use them to develop behaviour.

The phrase world model can sound larger than the systems it describes. In this research, it means a learned model of an environment’s dynamics, useful for predicting what may follow an action. Its value lies in what it enables an agent to do. A convincing picture of the future is not, by itself, proof that the model will support dependable decisions.

A path through learning and research

In his 2026 Innovators Under 35 profile, Hafner recalls discovering machine learning while attending high school in rural Germany, through Andrew Ng’s online course. The account places an accessible course near the beginning of a career that would later move between universities and industrial research. It is a useful human detail: before building agents that learn, he was finding his own route into the field.

His personal biography records an MRes at UCL and the Gatsby Unit, working with Tim Lillicrap and Karl Friston, followed by doctoral research at the University of Toronto with Jimmy Ba. He also spent time as a visiting student at UC Berkeley with Pieter Abbeel and undertook internships at Google Brain. Toronto’s repository dates his thesis, Embodied Intelligence Through World Models, to November 2024.

His industrial research career included a position as a Staff Research Scientist at Google DeepMind. By the time of his 2026 Innovators Under 35 recognition, the award profile described him as having left the company to establish a startup. His research can therefore be followed across institutions without treating an earlier affiliation as a permanent identity.

Planning from pixels

An important step was PlaNet, presented at ICML in 2019. With Tim Lillicrap, Ian Fischer, Ruben Villegas, David Ha, Honglak Lee and James Davidson, Hafner investigated learning dynamics directly from images. The agent performed its planning in a compact internal representation, rather than reasoning over every pixel as though it were an independent fact.

PlaNet connected perception with prospective action: learn how the environment changes, consider possible action sequences, then act. Its experiments addressed visual control tasks and compared learning efficiency and performance with other approaches. They did not amount to a general model of the physical world, but showed that a learned representation could be accurate enough to support useful planning in those settings.

Dreamer took the idea in a different direction. Developed with Lillicrap, Jimmy Ba and Mohammad Norouzi and presented at ICLR in 2020, it learned a behaviour policy from trajectories generated inside its model. The agent still needed experience to learn that model. The change was where much of the subsequent policy learning happened: imagined trajectories could provide training opportunities without requiring a fresh environmental interaction for each one.

When the learner has a physical body

DayDreamer brought the programme into physical robotics. Philipp Wu, Alejandro Escontrela and Hafner shared lead authorship, alongside Ken Goldberg and Pieter Abbeel. The CoRL 2022 work tested learning on four physical robots without first supplying an external simulator. One experiment reported a quadruped learning to walk from scratch in about an hour; other platforms examined manipulation and navigation.

The result concerned learning in particular experimental settings. It did not mean the robots required no training, nor that their internal models had disappeared. The distinction was between relying on a separately constructed simulator and learning a model from the robot’s own experience. Physical interaction supplied evidence that a purely visual demonstration could not.

A common method across different tasks

With Jurgis Pasukonis, Jimmy Ba and Tim Lillicrap, Hafner developed DreamerV3. The Nature paper, published in April 2025, reports results across more than 150 tasks using a common algorithm configuration. This is a claim about the breadth and robustness of a learning method, not a single trained agent that already possesses every tested skill.

The study includes collecting diamonds in Minecraft through learning from scratch, without human demonstrations or a manually organised curriculum. That qualification is part of the result. It should not be expanded into a claim that no previous AI system had ever collected a diamond, or that every later Dreamer experiment used the same training conditions.

The Dreamer 4 preprint of September 2025, co-led with Wilson Yan and co-authored by Timothy Lillicrap, studied training behaviour inside a model learned from recorded experience. Its Minecraft experiments used 2,500 hours of video with actions and event annotations. Evaluation also used task prompts. The authors reported diamonds in 0.7% of episodes: a notable but infrequent outcome, not reliable mastery. Learning without new interaction with the environment did not mean learning without human-produced data.

A career in seven moments

1. Education · Between universities and laboratoriesStudies at UCL and the Gatsby Unit; later works with research groups in Toronto and Berkeley.
2. 2019 · PlaNetCo-authors a method for planning from images through learned latent dynamics.
3. 2020 · DreamerDevelops policy learning through trajectories imagined by a world model.
4. 2022 · DayDreamerCo-leads experiments in learning with physical robots.
5. 2024 · A doctoral synthesisToronto records his thesis on embodied intelligence through world models.
6. 2025 · Broader control and offline learningDreamerV3 appears in Nature; the Dreamer 4 preprint explores a different learning setting.
7. 2026 · Recognition and a new institutional chapterAppears in MIT Technology Review’s Innovators Under 35 after leaving Google DeepMind.

Hafner’s trajectory makes a clear contribution to the discussion of AI beyond language generation. Intelligence is also tested in the passage from observation to prediction and from prediction to action. His work explores that passage through methods other researchers can examine and extend, while leaving an essential question open: how far can a model learned from limited experience be trusted when the situation changes?