Browser Runtime
Step a one-step Mario world model directly in the browser with learned action codes. This one is still a work in progress: the decoder is weak, the visuals are rough, and it probably suffers from having too many codes available relative to the quality of the learned semantics.
Past context from the dataset and the sequence you choose here.
The model predicts the next latent step and renders it directly as an RGB frame.
Choose the three-frame context used to start the rollout.
This Mario demo is much more forward-looking than the PushT one. The basic idea is already there, but the decoder is still weak enough that the model’s imagined futures look rough and unstable.
I also suspect the action space is over-fragmented here. With too many codes available and only a weak decoder to make their consequences legible, the semantic differences between buttons are harder to read than they should be.