cgearhart parent
The current gen VLA architectures include some tricks (like compressed action tokenization and diffusion decoding) to reach action frequencies between 50-200hz. I think they’re _more_ efficient this way than regular LLMs trying to do everything thru text.