World Action Model

A World Action Model is a way for a robot to understand the world not through photos and text, but by watching video and first picturing in its 'mind' how things will physically change when touched or moved — before it actually moves. It's like the model has already learned from video how a cup gets dented depending on finger shape when gripped, or how a towel folds when grabbed. Then, when the actual robot arm moves, it decides its next action based on this internal prediction.

Traditional robot 'brains' were built by taking a model trained to understand images and text together, and then bolting on an arm-control function afterward. These models are good at describing a scene in words, but weak at foreseeing how an object will actually deform once touched. A World Action Model swaps out that weakness for a prediction ability learned directly from video.

In the article, it appears like this: NVIDIA introduced the World Action Model (WAM) as a new foundation for robot policies. A common misunderstanding is treating this as the name of a single specific product, but it actually functions more as an umbrella term for an entire class of models that decide actions by predicting physical changes from video.