Thanks! We did indeed explore these additional degrees of freedom, you can find vestigial code for this here: https://github.com/google-research/robopianist/blob/main/rob...

We ended up picking a minimal subset of forearm DoFs that wouldn't impact training speed too much.