This work from Berkeley + Google Brain describes using RL and GANs for transferring motion styles from animals successfully onto robots.<p>Super neat idea, and love to see ML being used more and more on real robots!<p>Love the Cost of Transport analysis - naturalistic movement really is very efficient, and good luck getting RL to solve the hard exploration problem of good motion and task performance simultaneously tabula rasa!<p>It's great to see hand engineered reward functions go the way of the dodo[2] - people are finally coming around to the idea that it was just feature engineering in a difficult-to-reason-about (reward) space! Down with hand specified reward functions! Let imitating nature reign supreme.<p>What's next, GANs for moral style transfer?[1]<p>[1] <a href="https://imgflip.com/i/6asa47" rel="nofollow">https://imgflip.com/i/6asa47</a>