Register
World action models, classified
Each model is sorted by what it predicts. Inverse dynamics models generate the future video, then infer actions from it. Joint prediction models emit video and actions together. Representation-only models use the video backbone without generating frames. Latent-action models learn actions as hidden codes from video.
- BagelVLAA VLA that also predicts visual futures.
- BeingBeyond/Being-H0.7Latent world-action model bridging the two recipes.
- Sereact/Cortex 2.0Uses a world model for planning.
- NVIDIA/Cosmos PolicyEncodes actions as extra latent video frames.
- DreamDojoLearns latent actions from video.
- NVIDIA/DreamZero1,750 Elo on RoboArena (April 2026 snapshot), ahead of Pi-0.5 at 1,622. Trained only on DROID.
- Rhoda AI/DVADirect video-action model.
- 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.
- Ant Group/LingBot-VAMixture-of-transformers design; 16,000 hours of pretraining.
- Odyssey/Odyssey-3Foundation world model; Odyssey trains an action decoder on its representations. Research preview as Odyssey-3 Flash; no open weights stated.
- mimic-videoVideo-first policy that reads actions off generated futures.
- MotusMixes the VLA and WAM recipes.
- Unified Video ActionGenerates video and actions in one model.
- GENIMADraws the target actions into the image, so the action is itself a picture.
- Video Prediction PolicyUses predicted video representations to drive the policy.
- GR-13.06 of 5 subtasks on CALVIN ABC→D, when prior methods in its table stayed below 1.
- UniPiEarly recipe: generate the future video, then infer the actions that produce it.
Classification follows Moritz Reuss's field map on the NVIDIA Technical Blog (June 2026) and each model's own paper. A dash means we have not confirmed the value. Download the same rows as JSON or CSV.