Section
World models and video backbones
The pretrained models WAMs inherit from: Cosmos, Wan, V-JEPA, Genie and the video generators that learn how the world moves.
A 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. That usage follows Ha & Schmidhuber, World Models (2018) and the glossary in Reuss, NVIDIA Technical Blog (15 June 2026). A world action model is the overlap: a policy that reuses a world-model or video backbone and also emits actions.
Reuss splits the video side of that overlap into language-conditioned video generators (his examples include Cosmos, Wan, Veo) and action-conditioned world models (DreamDojo, Genie, JEPA-style latent dynamics). WAMs such as DreamZero, LingBot-VA, UniPi and mimic-video sit at the intersection. This section tracks the backbones those policies inherit, not a second full register of every video model.
Backbones currently named on the register, each confirmed from the model's paper or lab page: UniPi (Video diffusion (~1.7B)); GR-1 (GPT-style video pretraining); GENIMA (Stable Diffusion); Video Prediction Policy (Stable Video Diffusion); mimic-video (Cosmos-Predict2); Cosmos Policy (Cosmos); LingBot-VA (Wan 2.2 (5B)); DreamZero (Wan 2.1 (14B)); DreamDojo (Cosmos-Predict2.5); BagelVLA (BAGEL); Fast-WAM (Wan 2.2 (5B)); DVA (Causal video model). Rows with a dash have no confirmed backbone; we do not fill them from neighbouring papers.
Two families appear more than once. Wan 2.2 (5B) is listed for LingBot-VA and Fast-WAM. Wan 2.1 (14B) is listed only for DreamZero; the strings are kept distinct. Cosmos names are also kept exact: Cosmos Policy lists Cosmos, mimic-video lists Cosmos-Predict2, DreamDojo lists Cosmos-Predict2.5.
Open video weights are the practical reason WAMs moved after UniPi. Reuss's estimate of UniPi-style from-scratch CNN video pretraining is on the order of 1023 FLOPs; modern inverse-dynamics WAMs fine-tune open DiT backbones instead. That compute figure is Reuss's own rough estimate. Reuss, NVIDIA Technical Blog (15 June 2026)
Models in this section
- 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.
- mimic-videoVideo-first policy that reads actions off generated futures.
- 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.
- Rhoda AI/DVADirect video-action model.
- 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.
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.