# World Action Models > News and reference on world action models: robot policies built on pretrained video and world-model backbones. A static reference site. News and explainers are sourced from papers, lab pages and model cards. The register classifies world action models by paradigm; a dash means the value is unconfirmed. ## Pages - [Home](https://wamodels.com): News and reference on world action models: robot policies built on pretrained video and world-model backbones. - [News](https://wamodels.com/blog): Every news post and explainer on this site, each with a one-sentence summary. - [What is a world action model?](https://wamodels.com/posts/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. - [Register](https://wamodels.com/models): Every world action model we track, classified by paradigm, with backbone, lab and paper. - [Compare](https://wamodels.com/compare): Side-by-side comparison of register rows and sourced claims. - [Register JSON](https://wamodels.com/models.json): The register as JSON, one object per model from lib/models.ts. - [Register CSV](https://wamodels.com/models.csv): The register as CSV, same fields as the JSON dump. - [About](https://wamodels.com/about): Who publishes wamodels.com, editorial method, and sourcing rules. - [RSS](https://wamodels.com/feed.xml): Feed of news posts. - [llms.txt](https://wamodels.com/llms.txt): This file: a plain-text index of every public page. - [Sitemap](https://wamodels.com/sitemap.xml): Machine-readable list of every public page. ## News - [Odyssey-3 tests a video world model as a robot backbone](https://wamodels.com/posts/odyssey-3): 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](https://wamodels.com/posts/nvidia-field-map-world-action-models): 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](https://wamodels.com/posts/wam-speed-problem): 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?](https://wamodels.com/posts/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. ## Register - [UniPi](https://wamodels.com/models/unipi): inverse dynamics (2023). Early recipe: generate the future video, then infer the actions that produce it. - [GR-1](https://wamodels.com/models/gr-1): joint prediction (2023). 3.06 of 5 subtasks on CALVIN ABC→D, when prior methods in its table stayed below 1. - [GENIMA](https://wamodels.com/models/genima): joint prediction (2024). Draws the target actions into the image, so the action is itself a picture. - [Video Prediction Policy](https://wamodels.com/models/video-prediction-policy): inverse dynamics (2024). Uses predicted video representations to drive the policy. - [Unified Video Action](https://wamodels.com/models/unified-video-action): joint prediction (2025). Generates video and actions in one model. - [mimic-video](https://wamodels.com/models/mimic-video): inverse dynamics (2025). Video-first policy that reads actions off generated futures. - [Motus](https://wamodels.com/models/motus): hybrid (2025). Mixes the VLA and WAM recipes. - [Cosmos Policy](https://wamodels.com/models/cosmos-policy): joint prediction (NVIDIA, 2026). Encodes actions as extra latent video frames. - [LingBot-VA](https://wamodels.com/models/lingbot-va): inverse dynamics (Ant Group, 2026). Mixture-of-transformers design; 16,000 hours of pretraining. - [DreamZero](https://wamodels.com/models/dreamzero): joint prediction (NVIDIA, 2026). 1,750 Elo on RoboArena (April 2026 snapshot), ahead of Pi-0.5 at 1,622. Trained only on DROID. - [DreamDojo](https://wamodels.com/models/dreamdojo): latent actions (2026). Learns latent actions from video. - [BagelVLA](https://wamodels.com/models/bagelvla): hybrid (2026). A VLA that also predicts visual futures. - [Fast-WAM](https://wamodels.com/models/fast-wam): representation only (2026). Skips 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. - [Cortex 2.0](https://wamodels.com/models/cortex-2-0): hybrid (Sereact, 2026). Uses a world model for planning. - [DVA](https://wamodels.com/models/dva): inverse dynamics (Rhoda AI, 2026). Direct video-action model. - [Being-H0.7](https://wamodels.com/models/being-h0-7): latent actions (BeingBeyond, 2026). Latent world-action model bridging the two recipes. - [Odyssey-3](https://wamodels.com/models/odyssey-3): representation only (Odyssey, 2026). Foundation world model; Odyssey trains an action decoder on its representations. Research preview as Odyssey-3 Flash; no open weights stated. ## Sections - [World action models](https://wamodels.com/beats/wam): 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. - [Vision-language-action models](https://wamodels.com/beats/vla): The other recipe: policies built on vision-language models. Pi-0, GR00T and their successors, and the hybrids forming between the two. - [World models and video backbones](https://wamodels.com/beats/world-models): The pretrained models WAMs inherit from: Cosmos, Wan, V-JEPA, Genie and the video generators that learn how the world moves. - [Benchmarks and data](https://wamodels.com/beats/benchmarks): RoboArena, LIBERO, CALVIN, DROID and the other datasets and leaderboards that decide which recipe is winning.