It's honestly quite shocking how sparse the research and implementations for everything is once you go beyond a single sentence/command that you shout at your personal assistants.
Frames used to be an idea in AI, but they seem to have been sidelined and possibly forgotten now.
Frames mean that words and sentences have a context, and you can't understand conversations unless you understand the context.
This starts from simple and obvious distinctions. E.g. - as a silly example - "make dinner" usually means "Prepare and cook an evening meal". But if you have a project called "dinner" it might mean "build and compile 'dinner'" An AGI should be able to understand the difference, and ask for clarification if it doesn't.
Eventually you end up with subtextual and implied communication - e.g. "I'm fine" can mean two completely opposite things depending on tone of voice and the contents of minutes-to-years of previous conversations.
All of this is many orders of magnitude harder to handle than "Bedroom lights off."
Oh, I didn't mean AGI levels of understanding, but even "simple" technical things that are likely building blocks necessary to get to that point like sentence boundary detection.
That's fair, NLP has done decently with mechanical deconstruction of normal sentences for quite a while now. But as you note, mapping that onto a template for response is a long way from "understanding".
