World Action Models and Platforms

A world action model is a robot policy built on a pretrained video or world-model backbone: it predicts how the scene will change and outputs the actions to bring that change about. Vision-language-action models start from a vision-language model and map observations and instructions straight to actions, so grounding language in physical behaviour must be learned mostly from robot data. WAMs start from a video model that already represents how actions change a scene, on the argument that video priors can close that gap more easily. NVIDIA's overview treats WAMs as a possible second major recipe for robot foundation models next to VLAs, while noting that many claims rest on qualitative and simulation evidence.

NVIDIA groups WAMs into three formulations. Inverse dynamics: generate a future video, then infer the actions that would produce it (for example UniPi and LingBot-VA). Joint prediction: predict video and actions together (for example GR-1, DreamZero and Cosmos Policy). Representation-only: use the video model as a feature extractor and skip video generation at inference (for example Fast-WAM). Actions enter the model as ordinary tokens, as image-like targets the video model can denoise, or as latent plans, and Mixture-of-Transformers is currently the most common architecture.

Work in 2026 focuses on robustness, object structure and the high inference cost of generating video.

Model families, simulators and tooling that WAM work builds on.

General world models and company programmes that overlap with WAMs.

The established recipe that WAMs are compared with.

Evaluation of generalist robot policies.

WAMs usually need more training compute than VLAs because video sequences are around ten times longer, and they are slower at inference when they generate video. NVIDIA's blog quotes about 590 to 800 ms per action chunk for WAMs that generate video against about 190 ms for pi-0.5, and a separate 2026 paper reports 211.7 ms for Fast-WAM against 71.4 ms for pi-0.5; the set-ups differ, so the numbers are not comparable. Veo 3.1 test artifacts show that video models can still change a gripper's shape or morph the robot, so physical fidelity is limited. Robot data is scarce, large robotics-first datasets are mostly held by well-funded companies, and cross-paper comparisons are rough because backbones, data and evaluation differ.

Neighbouring pages on this site.