I tried refining Gr00t using the scripts provided in lerobot. Unfortunately my recorded data was with different input image resolutions. Not sure if you think it makes sense to support this, but would be great.

It then cancels in processor_groot.py, when around line 300 all the camera images are stacked.

I guess: aspect respecting rescaling, or just rescaling or a better way (perhaps it already exists) to control the input pipeline for video in training and inference would be great. I am probably a little bit ignorant here, did not write my own code just used the scripts provided by lerobot.