What is a world action model?
A robot policy that starts from a model pretrained to predict video, then learns to act. Here is how it works, how it differs from a VLA, and what it costs.
A world action model (WAM) is a robot policy built on top of a model that was first trained to predict video. The video model has learned, from huge numbers of clips, how scenes change over time: objects fall, doors swing, liquids pour. The WAM keeps that knowledge and is then fine-tuned on robot data so that it also outputs actions.
Moritz Reuss's field map of world action models on NVIDIA's developer blog puts it compactly: a WAM is a policy that starts from a pretrained world-model or video backbone and adapts it to predict how the scene will change.
Three terms that get confused
- World model. Predicts the next state of the world given an action. The state might be video frames, object positions or an internal representation. On its own it does not decide what to do.
- Vision-language-action model (VLA). A policy built on a vision-language model, the kind of model that describes images and answers questions about them. Physical Intelligence's π0 is the best-known example. VLAs inherit knowledge of objects and language, but must learn from robot data how a scene physically responds.
- World action model. A policy built on a world model or video model instead. It inherits knowledge of how things move, and must learn from robot data which actions cause which changes.
The difference is what each recipe gets for free. A VLA knows what a mug is. A WAM knows what a mug does when it is pushed.
How a WAM is trained
- Pretrain a video backbone, usually on very large text-conditioned video datasets. Open video models such as Wan are common starting points.
- Fine-tune on robot demonstrations so the model outputs actions.
- Keep predicting video while learning actions. DreamZero and Fast-WAM both report that WAMs work best when the video-prediction objective stays switched on during action training.
Three ways to get actions out
Reuss sorts WAMs by what they predict:
| Paradigm | What it does | Examples |
|---|---|---|
| Inverse dynamics | Generates the future video, then infers the actions that would produce it | UniPi, LingBot-VA, mimic-video |
| Joint prediction | Emits video and actions together | GR-1, DreamZero, Cosmos Policy |
| Representation only | Uses the video backbone's internal features and skips generating frames | Fast-WAM |
Actions can also be represented differently: as ordinary action tokens, drawn into images (GENIMA, Cosmos Policy), or as latent codes learned from unlabelled video (Genie, LAPA). The register classifies every model we track.
Why people are excited
On RoboArena, a benchmark where robot policies are compared head to head and ranked by Elo, NVIDIA's DreamZero scored 1,750 in the April 2026 snapshot Reuss cites. Physical Intelligence's π0.5 scored 1,622. DreamZero was trained only on the DROID dataset.
What it costs
Generating video is slow. Fast-WAM's Fig. 4, measured on one real towel-folding task on a single NVIDIA RTX 5090D V2 32GB GPU, puts Fast-WAM-Joint at 580 milliseconds and Fast-WAM-IDM at 810 milliseconds per chunk of actions. Those are Fast-WAM's own variants, not WAMs in general. π0.5 is 180 ms on the same plot; Fast-WAM itself is 190 ms. That is about three to four and a half times slower than π0.5 for the imagine-then-execute variants, which matters for a robot that has to react.
Training is also expensive. Reuss estimates, by his own description roughly, that a full WAM built on a 14-billion-parameter video model takes around seven times the compute of an equivalent VLA path. Representation-only designs such as Fast-WAM exist largely to claw back that speed.
Where it is heading
Reuss expects WAMs to become the second main recipe for robot foundation models, and the eventual winner to be a hybrid. Models such as Motus, BagelVLA and Being-H0.7 already mix the two. Odyssey-3, a closed research preview, makes the same bet from the other side: a video world model plus a small action decoder as the thing a robot acts from. This site tracks that contest.