You can think of these as essentially multi-modal LLMs, which is to say you can have very small/fast ones (SmolVLA - 0.5B params) that are good at specific tasks, and larger/slower more general ones (OpenVLA - a finetuned llama2 7B). So a rpi could be used for some very specific tasks, but even the more general ones could run on beefy consumer hardware.

According to the blog post, it requires an NVIDIA Jetson Orin with at least 8GB RAM, and they've optimized for Jetson AGX Orin (64GB) and Orin NX (16GB) modules.

Could you quote where in the blog post they claim that? CTRL+F "Jetson" gave no results in TFA.

Yeah they didn't really mention anything, I was almost getting my hopes up that Google might be announcing a modernized Coral TPU for the transformer age, but I guess not. It's probably all just API calls to their TPUv6 data centers lmao.

There's a post on x from one of the project contributors that says it fits on a 4090: https://x.com/sippeyxp/status/1937520297789497668