Huge Step Forward in Legged Robotics from ETH

Control policies learned via RL are starting to work in the real world!

Typically policies learned via simulation tend to transfer poorly to the real world (the so-called sim2real gap), so I'm curious to dig into this work to see how they overcame this limitation.

From just watching the video and guessing, it would make sense if noising the belief state (rnn(h,concat(proprio,extero)) + \eps ~ Noise) and learning to condition proprioceptive attention on the belief uncertainty is enough. Very cool work and so exciting to see robotics groups exploiting ML more and more (gate attention + learned belief states here).