Section
World action models
Policies that start from a video or world-model backbone and learn to predict both how a scene changes and what the robot should do.
A world action model is a robot policy built on a backbone that was first trained to predict video or future world state, then adapted so it also outputs actions. That definition is the one this site uses in what is a world action model?, following Reuss, NVIDIA Technical Blog (15 June 2026).
Reuss sorts WAMs by what they predict. Inverse dynamics: Generates the future video, then infers the actions that would produce it. On this register: UniPi, Video Prediction Policy, mimic-video, LingBot-VA, DVA. Joint prediction: Emits video and actions together. On this register: GR-1, GENIMA, Unified Video Action, Cosmos Policy, DreamZero. Representation-only: Uses the video backbone's internal features and skips generating frames. On this register: Fast-WAM, Odyssey-3. Latent actions: Learns actions as hidden codes from video. On this register: DreamDojo, Being-H0.7. Hybrid: Mixes the vision-language-action (VLA) and world action model recipes. On this register: Motus, BagelVLA, Cortex 2.0.
Those rows are the 17 models in the register. Each card links claims to the paper, project page or official repo they came from. A dash means the field is unconfirmed. Classification follows Reuss's three axes — what the model predicts, how actions enter, and how the parts are wired — checked against the model's own source.
Two results from that map are worth keeping attached to their sources. On RoboArena, a distributed real-world comparison of generalist policies (Atreya et al., RoboArena (2025)), Reuss's April 2026 snapshot puts NVIDIA's DreamZero at 1,750 Elo against 1,622 for Physical Intelligence's π0.5. Fast-WAM is the register's representation-only row: its paper reports 190 ms latency by skipping video generation at test time. Fig. 4, on one real towel-folding task on an RTX 5090D, puts Fast-WAM-Joint at 580 ms, Fast-WAM-IDM at 810 ms, and π0.5 at 180 ms. Those 580 / 810 ms figures are Fast-WAM's own variants, not WAMs in general.
Reuss expects WAMs to become the second main recipe for robot foundation models, and the eventual winner to be a hybrid. This section holds news filed under that recipe. The register is the backlink engine; this page is the stream.
On the register
- UniPiEarly recipe: generate the future video, then infer the actions that produce it.
- GR-13.06 of 5 subtasks on CALVIN ABC→D, when prior methods in its table stayed below 1.
- GENIMADraws the target actions into the image, so the action is itself a picture.
- Video Prediction PolicyUses predicted video representations to drive the policy.
- Unified Video ActionGenerates video and actions in one model.
- mimic-videoVideo-first policy that reads actions off generated futures.
- MotusMixes the VLA and WAM recipes.
- NVIDIA/Cosmos PolicyEncodes actions as extra latent video frames.
- Ant Group/LingBot-VAMixture-of-transformers design; 16,000 hours of pretraining.
- NVIDIA/DreamZero1,750 Elo on RoboArena (April 2026 snapshot), ahead of Pi-0.5 at 1,622. Trained only on DROID.
- DreamDojoLearns latent actions from video.
- BagelVLAA VLA that also predicts visual futures.
- Fast-WAMSkips video generation at inference. Fig. 4 on one RTX 5090D towel-fold: Fast-WAM 190 ms, Fast-WAM-Joint 580 ms, Fast-WAM-IDM 810 ms, Pi-0.5 180 ms.
- Sereact/Cortex 2.0Uses a world model for planning.
- Rhoda AI/DVADirect video-action model.
- BeingBeyond/Being-H0.7Latent world-action model bridging the two recipes.
- Odyssey/Odyssey-3Foundation world model; Odyssey trains an action decoder on its representations. Research preview as Odyssey-3 Flash; no open weights stated.
Stories
Odyssey-3 tests a video world model as a robot backbone
Odyssey-3 is a video world model that Odyssey says can also run robot arms, humanoids and cars. We check the claimed robot results against Physics-IQ Verified.
NVIDIA researcher maps world action models as robotics' second recipe
Moritz Reuss's long post sorts more than a dozen models into three paradigms, and predicts the field will end up with VLA–WAM hybrids.
The speed problem: why video-generating world action models are slower than VLAs
Imagining the future in pixels before acting costs hundreds of milliseconds. Representation-only designs try to keep the knowledge and skip the rendering.
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.