I do like that high quality (« natural ») motions for humanoids are now very much appreciated. Welcome to the first steps beyond the « uncanny valley »!!

I’m a bit less impressed, except for the hardware (great!) as we knew how to do the control as of 2013, 2017, 2018, depending on one’s point of reference. That sim-to-real now works relatively well ( with a bit of effort) is perhaps the biggest surprise over the past 6-8 years.

Vladlen Koltun’s paper on alternatives to h-indices is relevant and worth a read: vladlen.info/publications...

Deepmimic++ , i.e., Mocap tracking with RL, with appropriate retargeting and some env randomization?

Free energy or bust! 😆

Our first rule is: everyone shall ask questions, all the time, no matter how basic. Anyone can jump in to answer any question. Someone who hasn’t read a paper can add surprising value in many ways! Corollary: if there are no questions, no one is understanding anything.

If I have time, I'll directly time-index to the various talks. I'm also happy to share this work with others :-)

Yes, I had a similar experience! Amazing learners!