odf posted at 5:47 PM Mon, 1 November 2021 - #4429733

I think machine learning would probably be overkill here. Modulo any potentially tricky details, the rough outline of the algorithm would be to make a pose for Antonia that matches the donor figure, convert that and the donor pose to general matrix form (or quaternions if one feels so inclined) compose the matrices and convert back to Euler angles. Not entirely unlike IK, come to think of it, but simpler.

I was just joking, pointing out the vast resources available. One may get there, though, at some point as a part of a higher goal of natural figure control: user giving giving orders and the figure carrying them out in a natural way. But, yes, alas we're still far away from there. 

The approach you indicate for pose conversion is indeed the most logical route when only the end pose matters, as is the case when using conforming clothing. For dynamic clothing the route to the final pose is important and that is where the AI approach could come in because the feasible space of intermediate poses is quite curved and complex and the figure has to navigate those narrow channels. Trial and error would work but I expect this is impractical and time-consuming.

Quaternions or matrices is a matter of preference. Quaternions strike as more robust than Euler matrices because they do not have the gimbal lock and therefore also in general have less risk of bad condition. 

What does really help is to have a good library of hand poses. Fingers are the end of the chain and that is where converted poses need the most work, especially when they need to connect to something. Often less work to just reset the hands and start with target figure's hand dials or presets than sort out the mess. Maybe a pose converter could just ignore fingers initially.