Most video-action robot models are a video generator built for content， with an action head bolted on.

LingBot-VA 2.0 pretrains the whole stack for control from scratch.

Lots of technical improvements here that stand out：