If they've managed to embed the video collector at eye height or close, so the effect of looking at the screen is to be seen as looking at the other person, this is in itself useful. If it's still a non-eyeline camera and they do some adaptive smarts to make it "look" like eyeline I'd be interested how good that is.

If they have leveraged peppers ghost and like techniques to make a volume appear to be "occupied" then it's possible this also helps with hinting to presence. I've seen some art displays which sort-of did this. It was remarkably pleasant to sit opposite people who you sort-of knew were not actually in the implied volume behind the screen.

The influence of delay on perception of "there-ness" is huge. I would be interested how good this is at adjusting for unremovable lightspeed delay inter-continental. Not to over do it, it wouldn't surprise me if this works in SFO to NYC but works less well SFO to LHR or SFO to SIN.

I tend to assume "applied ML" is marketing speak but its possible the investment in time to tune some net has made an FPGA cheap to deploy to do one job, and do it well, and over time be (re)tuned to do it better. This isn't welcome to new robot overlords, its just sensible use of technology to improve.

I could see anything which does image processing like this also be applicable for people with persisting tic or tremor issues. It might help with stabilising image for their remote participation in online stuff.

I totally prefer this kind of application of smarts, to what Meta are doing.