The speed problem: why video-generating world action models are slower than VLAs
Imagining the future in pixels before acting costs hundreds of milliseconds. Representation-only designs try to keep the knowledge and skip the rendering.
A world action model that generates video before it acts has to render the future before it can do anything. That is expensive. Fast-WAM's Fig. 4, measured on one real towel-folding task on a single NVIDIA RTX 5090D V2 32GB GPU, puts Fast-WAM-Joint at 580 ms and Fast-WAM-IDM at 810 ms per action chunk. Those are Fast-WAM's own imagine-then-execute variants, not WAMs in general. Physical Intelligence's π0.5, a vision-language-action model, is 180 ms on the same plot; Fast-WAM itself is 190 ms.
For a robot, a few hundred milliseconds is the difference between catching a falling cup and watching it fall. Robots get around this partly by predicting a chunk of several future actions at once, but the gap still limits how reactive a WAM can be.
Skip the pictures
Fast-WAM's answer is to keep the video backbone and stop generating video at inference time. The model uses the backbone's internal representation, which already encodes how the scene is likely to evolve, and decodes actions from that directly. Reuss expects these representation-only designs to become a larger research area.
Train with video, act without it
There is a subtlety. DreamZero and Fast-WAM both find that WAMs work best when they keep a video-prediction objective during training. The pictures appear to matter as a teacher even if the deployed robot never draws them. That suggests the speed problem is an inference problem, not a reason to abandon video pretraining.
We will track latency figures in the register as more models report them.