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.
1. Well, based on Karpathy's talks on Tesla FSD, his solution is to actually make the training set reflect everything you'd see in reality. The tricky part is that if something occurs 0.0000001% IRL and something else occurs 50% of the time, they both need to make 5% of the training corpus. The thing with multimodal LLMs is that lidar/depth input can just be another input that gets encoded along with everything else, so for driving "there's a blob I don't quite recognize" is still a blob you have to drive around.
2. Figure has a dual-model architecture which makes a lot of sense: A 7B model that does higher-level planning and control and a runs at 8Hz, and a tiny 0.08B model that runs at 200Hz and does the minute control outputs. https://www.figure.ai/news/helix
1) Properly recognize what they are seeing without having to lean so hard on their training data. Go photoshop a picture of a cat and give it a 5th leg coming out of it's stomach. No LLM will be able to properly count the cat's legs (they will keep saying 4 legs no matter how many times you insist they recount).
2.) Be extremely fast at outputting tokens. I don't know where the threshold is, but its probably going to be a non-thinking model (at first) and probably need something like Cerebras or diffusion architecture to get there.