[saint_angels]

I think the main problem here is with transitions between animation states and matching each animation to the context. If ML motion matching[1] becomes popular in the industry, I think it would smooth out these issues a lot.

[1]https://www.youtube.com/watch?v=16CHDQK4W5k

[ricardobeat]

I remember this being solved over ten years ago, without ML, there was this software that seamlessly blended between poses and activities (morphine or something like that?). I wonder what happened to them. Unreal 5 also has similar tech.

[saint_angels]

can't find anything on "morphine", but animation blending is something that Unity/Unreal/etc had for a while. Although, I don't think any publicly available engine solves this like Ubisoft ML models do. Usually it requires at lot of work to look seamless

[ricardobeat]

The demo was a blender-like environment, with opaque white human models. You could shoot arrows at them, bullets, switch between poses and animations or set up obstacles to watch them react realistically and tumble around.

This was sometime before 2010.