HP did this at least 15 years ago with their "Halo" conferencing systems:
https://www.networkworld.com/article/2258553/inside-the-halo...
https://www.networkworld.com/article/2258553/inside-the-halo...
From watching the video, Google's conferencing setup is creating a 3D rep of the people talking and adjusting rendered view based on where the participants are seated. This is blending AR with videoconferencing. It would be interesting to see how their conferencing system works with multiple-people on each side. I know the video had a mother and baby in one segment, however is the 3D rendering based on the eye-level of the main participants?
Potentially a solved problem, just fix it in post. :)
https://www.zdnet.com/article/windows-10-microsoft-finally-l...
I wonder how well it works, and how much latency (if any) it adds to the feed.
I'm also sad it isn't rolled out more generally, a very strange feature to lock behind a small-ish volume hardware product.