About the Inference Pipeline

[heyicheng-never]

Thank you for your work and for open-sourcing it!

I have a question about how the model is applied in inference. According to the pipeline, the Gaussian World Model requires a^(t) as input to predict the next GS, while the model ultimately predicts a^(t) through PerceiverIO. This means that the model’s prediction is used as an input for the model itself, which confuses me.