Why are register tokens disabled during pre-training but enabled during fine-tuning?

[einnullnull]

Hi, and thank you for open-sourcing this excellent project!
I noticed that num_registers = 4 is used only during fine-tuning. Was this setting shown to be necessary—i.e., does removing registers cause a clear performance drop—or was it chosen mainly for convenience?
Have you run an ablation without registers in the fine-tuning stage?

[HongzheBi]

This is just a habit of mine. You can experiment yourself to see how much it specifically affects; I think the impact is very minor.

[einnullnull]

Thank you for your prompt response!

I have another question regarding the architecture. I noticed that you use a separate encoder for the state. For a given sample, the state seems to be the current qpos, while the action represents the future time step's qpos. Given this relationship, why did you choose to use a separate encoder for the state instead of sharing the encoder with the action? Is there a specific reason for keeping them distinct?

[HongzheBi]

Also a habit to prevent reciting the state.

[einnullnull]

Thank you for the explanation!