Gary Marcus

if they showed that in some general way that wasn't one (pre-rigged) object, I could maybe sorta squint and see that point. but sim to real with one overlearned prerigged object is a joke. no respectable academic journal would buy that argument.

They do have unseen modifications e.g. finger tying. While it's just one object, I really hope you appreciate how hard this must have been to pull off. While I'm with you that the PR is a bit silly, this is a significant advance.

Indeed your last sentence is not far my original assessment, repeated above, that "[despite the other misrepresentations] system is (very impressively) trained to do the perception and manual dexterity parts." I do remain skeptical that demo w 1 prerigged object solves sim->real.

There is a big gap between "solving sim->real" and "joke". Did anyone actually claim the former? It's an advance in sim2real capabilities that is significant while falling short of the hype.

can you really infer that from this demo?

Nope, but from the paper, yes.

any predictions of generalizability ito novel, uninstrumented objects you would like to put money on? & why so confident?

They definitely didn't solve that problem, but did make progress on a much simpler problem that hadn't been previously solved. I'm confident its an advance because the bar is *so* low. Getting a shadow hand to do anything is difficult, as is closing the simulation gap.